Skip to content
Corporate & Investment Banking My progress Français

Home / Capital Markets & Securities Services / Cross-functional / Module 27

27Counterparty Risk

Limits, CVA and the XVA complex — how a bank measures, prices, collateralises and closes out the risk that its client does not pay

Measuring and pricing the risk that a client does not pay.

Reading time
5 h 30
Sections
57
Recall questions
53
Quiz questions
15
Simulators
2

Two threads run through this module

The business thread is the story of a credit judgement turning into a price. A credit officer forms a view about a company — its leverage, its cash generation, its sector, its rating equivalent — and that view has to end up as a number of basis points on a trade ticket, quoted to a salesperson in the time it takes a client to stay on the phone. Along the way it passes through a limit framework that says how much of this counterparty the bank wants at all, a legal framework that determines whether offsetting trades can be set against each other in a bankruptcy, and a collateral framework that decides how much of the exposure is neutralised daily in cash. When the charge is too low the bank has sold protection it did not intend to sell. When it is too high the bank loses the trade to a competitor who understood the netting set better. Neither outcome is a modelling problem in the narrow sense; both are failures of judgement made visible by arithmetic.

The technology thread is the story of three systems that must agree with each other every single night. A simulation grid generates thousands of forward paths of rates, FX and credit, revalues an entire derivatives portfolio along each of them, and aggregates the results into netting sets — a computation that must finish between the market close and the start of the next trading day. A limit system takes those numbers and decides, in milliseconds the following morning, whether the next trade can be done. A collateral machine independently computes what should be called from each counterparty, reconciles it against what the counterparty's own system computed, and disputes the difference. The three run on different data, different release cycles and often different decades of technology, and the moment they disagree the bank is either blind or wrong. Most of the operational history of this business is the history of making them agree.

How to read this module

The module is written to be read in order, because each part assumes the vocabulary established by the one before it. Part 1 builds counterparty risk from first principles and places it in the bank and in the market. Part 2 walks the ten stages of the lifecycle, from onboarding to close-out, and is the spine of the document. Part 3 maps the actors and what each one controls. Part 4 goes inside the systems and the data. Part 5 draws out the risks, the real cases and the transformation agenda. A single running example — a Swiss chemicals company and a ten-year cross-currency swap — is carried through all five parts so that the same trade can be seen from the credit officer's desk, the simulation grid, the collateral operations team and the workout lawyer.

Seven kinds of box interrupt the prose, each with a fixed job. A Key insight box states the one thing to remember from a section, usually a claim that reframes what came before it. A Going deeper box holds the technical layer that a specialist will want and a generalist can skip without losing the thread. An IT & data lens box translates a business concept into what it means for systems, data models and interfaces. A Watch out box flags a common misunderstanding or a place where real institutions have got it wrong. A Worked example box carries the running case and every illustrative number attached to it. An In a nutshell box closes a part by compressing it to its load-bearing statements. A Check your understanding box ends each part with questions and answers, and is meant to be used rather than read.

What you will be able to do

You will understand the counterparty-risk lifecycle end to end, from onboarding to close-out

The lifecycle has ten stages, from the first credit file on a prospective client to the final settlement of a close-out amount years after a default. Each stage has an owner, a typical duration, a decision that gets made and a system where the decision is recorded. You will be able to say what happens at each stage, who acts, and what breaks when the stage is done badly. That end-to-end view is what lets you tell a modelling problem apart from a documentation problem apart from an operations problem.

You will be able to map every actor and say exactly what each one controls

Counterparty risk is unusual in how many functions have to touch a single decision. Coverage bankers and trading desks originate, credit risk approves, the XVA desk prices and hedges, Treasury funds, Collateral Operations calls and disputes, Finance reports the capital, Legal writes the agreement, and outside the bank there are clearing houses, custodians, data vendors and regulators. You will be able to draw that map, identify who has veto power at each point, and recognise the friction points where two functions share responsibility for the same number.

You will understand where the CVA charge actually comes from

CVA is not a fee and not a buffer. It is the market value of a contingent claim — the expected loss from a counterparty defaulting, computed as an expected exposure profile multiplied by default probability and loss given default, discounted back. You will be able to explain each of those three ingredients, where they come from, and why the answer moves when nobody has defaulted and nothing about the trade has changed. You will also understand why CVA sits alongside FVA, MVA, KVA and the rest of the XVA family, and what each of those is compensating the bank for.

You will be able to trace an exposure number front to back

A single figure on a limit report — say a peak future exposure against a group limit — is the output of a chain that starts with a trade capture message, passes through a market data close, a curve build, a Monte Carlo simulation, a netting-set aggregation and a collateral adjustment, and ends in four downstream consumers. You will be able to follow that chain, name the systems and the data at each hop, and identify where a number can silently go wrong. This is the skill that separates people who can discuss counterparty risk from people who can fix it.

You will recognise the risk, capital and transformation landscape

You will know the six risk types that a counterparty-risk function actually manages, and be able to attribute real failures — Archegos, Lehman, the 2022 margin crisis — to specific stages of the lifecycle rather than to bad luck. You will understand how SA-CCR, the internal model method, FRTB-CVA and the Basel 3.1 output floor determine what the business costs in capital. And you will be able to judge which transformation initiatives move an exposure or a capital number and which ones only move a dashboard.

1Counterparty risk is the price of being owed money in the future

Every derivative is a promise to exchange cash on dates that have not arrived yet. The moment a bank enters one, it acquires something it did not ask for and is not automatically paid for: a claim on a company that may not exist when the payment falls due. This part builds that idea from nothing, separates it cleanly from the credit risk on a loan and from the settlement risk in a payment, places the function inside the bank, ranks the products by how hard they are to manage, introduces the running example, sizes the market with real figures, and explains why the investment cycle in this business follows the rulebook rather than the market.

Counterparty risk is what is left when you strip a contract down to a promise

Start with a contract that pays you later and a party that might not be there later — that is the whole of it. Suppose a bank agrees today to pay a company a fixed rate on EUR 100m for five years, and to receive a floating rate in return. No money changes hands at the start. Every quarter, one side owes the other the net difference between the two rates. If rates fall, the floating leg the bank receives shrinks, the fixed leg it pays does not, and the contract becomes a liability to the bank. If rates rise, the contract becomes an asset. The bank's counterparty risk is the possibility that, on a day when the contract is an asset, the company on the other side stops paying.

The word "exposure" in this business means one specific thing: the amount you would lose today if your counterparty vanished today. If the contract is worth plus EUR 4m to the bank and the counterparty defaults, the bank has to go into the market and replace the trade at current prices, which costs it EUR 4m, and it joins the queue of unsecured creditors to recover what it can. If the contract is worth minus EUR 4m to the bank, the counterparty's default costs the bank nothing in replacement terms — the bank simply owes EUR 4m to the estate and pays it. Exposure is therefore the positive part of the mark-to-market and nothing else. Negative marks do not become negative exposures; they become zero. That one asymmetry is the source of almost every complication that follows.

Three properties make counterparty risk a different animal from the credit risk a bank has been managing since it was invented. The first is that the exposure is stochastic. Nobody knows what a ten-year swap will be worth in year six; it depends on where rates and FX go, so the exposure is a probability distribution over every future date, not a number. The second is that it is bilateral. Either party can end up owing the other, which means both sides are simultaneously each other's credit risk, and the contract can flip sign many times over its life. The third is that it is contingent on the same markets that drive the counterparty's own health. An oil producer that hedges by selling forward is a good credit when oil is high, but that is precisely when the hedge is worth money to the producer and nothing to the bank. When oil collapses, the bank is in the money on the hedge and the producer is in trouble — the exposure grows exactly as the ability to pay shrinks. That correlation has a name, wrong-way risk, and it is the reason counterparty risk cannot be handled by simply bolting a probability of default onto a market-risk system.

The exposure is a distribution, so every control is a choice of which point on it to act on

Because the future value of a derivative is unknown, the honest answer to "what is our exposure to this client?" is a picture rather than a figure. For each future date, there is a whole range of values the portfolio might take, with probabilities attached. The bank collapses that picture into summary measures, and the choice of measure is a decision about what the number is for. An average is the right input to a price, because a price should reflect what the bank expects to lose. A high percentile is the right input to a limit, because a limit exists to defend against a bad outcome rather than a typical one. Using the wrong one is not a rounding error; it is a category mistake that either overcharges every client or leaves the bank undefended in the tail.

A small vocabulary carries the whole of the rest of this module, and it is worth fixing now. Expected exposure, EE, is the average of the positive exposure on a given future date. Expected positive exposure, EPE, is the average of EE across time, and it is the quantity that feeds the CVA calculation. Effective EPE, EEPE, is a regulatory variant that prevents the measure from falling as short-dated trades roll off, and it is the input to capital under the internal model method. Potential future exposure, PFE, is a high percentile of the exposure distribution — commonly 95% or 97.5% — and it is what limits are set against. The margin period of risk, MPoR, is the assumed number of days between the last effective collateral exchange and the completion of close-out, and it determines how much exposure can accumulate on a collateralised portfolio despite the collateral.

Measure What it answers Where it is used
EE What do I expect to be owed on a given future date? Building the exposure profile
EPE What do I expect to be owed on average over the life? The CVA calculation
EEPE The same, with a regulatory non-decreasing adjustment Capital under the internal model method
PFE What is the worst I could plausibly be owed? Limit setting and the pre-deal check
MPoR How long am I exposed after collateral stops arriving? Collateralised exposure and IM sizing

These measures are computed rather than observed, and that has a governance consequence. No market quotes a bank's EPE. It comes out of a simulation whose assumptions — the volatilities, the correlations, the number of paths, the collateral logic applied along each path — are the bank's own choices, validated internally and, where an internal model is used for capital, approved by a supervisor. Part 2 shows how the numbers are produced and Part 4 shows which systems produce them. The point to carry forward is only that every figure on a counterparty-risk report is the output of a model, and the interesting question about any of them is always which point on the distribution it represents.

Credit risk on a loan is the same idea with the hard part removed

A loan has one exposure and it is known in advance; a derivative has a different exposure every day for the next ten years. When a bank lends EUR 100m it knows what it stands to lose from the moment the money leaves the building. The amount amortises on a published schedule, it never goes negative, and the borrower is unambiguously the debtor for the entire life of the facility. The bank prices this with a credit margin over funding, books it in the banking book, provisions against it, and knows within a rounding error what it is carrying. The exposure is deterministic and one-directional.

Take those two properties away and every control the bank uses has to be reinvented. A limit cannot be "EUR 100m drawn" because there is no drawn amount; it has to be a statement about a distribution, such as the 97.5th percentile of possible future exposure. A provision cannot be a percentage of a balance because there is no balance; it has to be a valuation of a contingent claim, which is to say a derivative price. Collateral cannot be a mortgage taken once at the start; it has to be recalculated and moved every day because the amount at risk changes every day. And recovery cannot be assessed loan by loan because a defaulting counterparty may have four hundred trades with the bank, some in the money and some out, whose treatment in insolvency depends on a contract signed years earlier.

Settlement risk is short, violent and conceptually separate

Settlement risk is the risk that you pay and the other side does not, inside the plumbing of a single transaction. It has nothing to do with distributions or ten-year horizons. It lasts hours, sometimes minutes, and its magnitude is the full principal amount rather than a replacement cost. In an FX trade, one currency settles in one time zone and the other settles in another; between the two moments, one party has irrevocably delivered and the other has not yet.

The canonical case is Bankhaus Herstatt, and it is worth knowing because it named the risk. On 26 June 1974, German regulators withdrew the banking licence of Bankhaus Herstatt, a mid-sized Cologne bank that had taken large speculative FX positions, and ordered it into liquidation during the German banking day. Counterparties had already paid Deutsche Marks to Herstatt that morning in expectation of receiving US dollars in New York later the same day. The Marks were gone; the dollars never came. The losses cascaded through the correspondent banking system and froze parts of it for days. The episode is why the risk is still called Herstatt risk, why the Basel Committee itself was created in the aftermath, and why the industry eventually built CLS — a settlement system that exchanges the two legs of an FX trade simultaneously so that neither side can be delivered without the other.

The distinction that matters is exposure size against exposure duration. Settlement risk is full-notional for a very short window. Counterparty credit risk is a fraction of notional for a very long window. They are managed by different mechanisms — payment-versus-payment infrastructure and daily settlement limits for the first, simulation, collateral and CVA for the second — and confusing them leads to real errors, most commonly the assumption that a spot FX book carries no counterparty risk when in fact it carries a large and very concentrated one for a few hours a day.

The function sits in the second line, originates in the first, and has one strange hybrid in the middle

Counterparty risk is governed under the three lines of defence, and the second line owns the veto. The first line — coverage bankers, sales and trading desks — originates the risk. They know the client, they see the flow, and they are compensated on revenue. The second line is independent credit risk: credit officers who write the counterparty analysis, assign the internal rating, set the limits and approve or refuse the trades that breach them. Their independence is structural, not cultural; they report up a chain that does not end at the head of markets, precisely so that a large ticket cannot be approved by the person who benefits from it. The third line, internal audit, tests whether the first two did what the policy says they do.

The credit officer's product is not an opinion, it is a set of numbers other systems will enforce. An internal rating that maps to a probability of default. A set of limits by tenor and by measure. A list of permitted products for that counterparty. A collateral requirement, if any. Once written, those numbers are consumed automatically — the pre-deal check that runs in the milliseconds before a trader quotes a price is reading the credit officer's decision from a database, not calling them on the phone.

The independence of the second line is structural rather than a matter of professional culture, and the design is deliberate. A credit officer's reporting line runs to a Chief Risk Officer, not to a head of markets, and their compensation is not tied to the revenue of the desks whose trades they approve. Approval authority is graded by size and by rating, so a small ticket with an investment-grade name is delegated to a single officer while a large one with a weak name escalates through a credit committee, and above a threshold it reaches a group-level body that also owns the appetite for the sector and the country. The escalation ladder is not bureaucracy for its own sake. It ensures that the decisions with the potential to hurt the institution are taken by people whose careers are not improved by taking them.

The friction this produces is real and is a feature. A salesperson with a client on the phone and a competitor quoting experiences the pre-deal check as an obstacle, and sometimes it is one — a limit set a year ago against a portfolio that has since changed shape, or a rating that has not been refreshed since the counterparty's last results. Every bank therefore runs an override and temporary-increase process, and every bank's supervisor looks closely at how often it is used and by whom. A pattern of repeated temporary excesses on the same counterparty is one of the most reliable early indicators that a limit framework has stopped reflecting the bank's actual appetite, and it appears in the post-mortem of more than one of the failures examined in Part 5.

The XVA desk is a trading desk that owns a risk-management output, and that is deliberate

Someone has to own the profit and loss of the CVA reserve, and it cannot be the sales desk. When a bank prices a derivative to an uncollateralised client, it adds a charge for the counterparty risk it is about to take. That charge is collected once, at inception, from the client. The risk, however, lasts a decade and changes value every day with rates, FX and the client's credit spread. If the money simply disappeared into the sales desk's revenue line, nobody would be managing the resulting position, and the bank would discover the problem only when a spread blew out and a reserve had to be topped up out of nowhere.

So the bank creates an internal market. The XVA desk buys the counterparty risk from the originating desk at the CVA price, takes it onto its own books, and thereafter manages it as a portfolio. The trading desk gets a clean trade, hedged for market risk, with the credit component stripped out — which is what a trading desk is equipped to manage. The XVA desk gets a book of counterparty credit exposure across the whole institution, which it hedges with single-name credit default swaps where they exist, index CDS where they do not, and interest rate and FX hedges for the sensitivity of the exposure profile itself. It reports a P&L, it has risk limits, and it is judged on whether the charges it collected covered the losses and hedging costs it incurred.

The hybrid nature is the point and also the permanent tension. The XVA desk is a first-line trading unit by structure and a risk-management function by purpose. It is the only desk in the bank whose inventory was created by other people's decisions, and it has no ability to refuse a trade — that veto belongs to credit risk. What it can do is price, and if it prices high enough the trade does not happen. In practice that makes the XVA desk the mechanism by which a soft credit view becomes a hard commercial constraint.

Four other functions hold pieces of the same problem

Treasury funds what collateral does not cover. An uncollateralised derivative that is in the money is an unsecured asset the bank must fund; a collateralised one requires the bank to post cash it must borrow. The cost of that funding is FVA, and Treasury sets the internal rate at which it is charged. Collateral Operations runs the daily margin cycle — computing calls, issuing and receiving them, settling movements, applying haircuts, and disputing differences with counterparties whose systems produced a different number. Finance and Capital Management turn the exposures into regulatory capital under SA-CCR or the internal model method, produce the CVA capital charge, and report the whole thing to supervisors. Legal negotiates the ISDA Master Agreement, the Schedule and the Credit Support Annex, and commissions the netting opinions that determine whether any of the offsetting the bank relies on will actually hold in a foreign bankruptcy court. Part 3 maps these actors in detail, including who has veto power at each decision point; the purpose here is only to place them.

The products form a ladder, and the difficulty rises faster than the complexity

Counterparty risk exists wherever there is a promise to exchange value later, but the effort required to manage it varies by orders of magnitude across products. The useful way to organise this is a ladder from simplest to hardest, because each rung adds exactly one new difficulty to the one below it. What changes as you climb is not primarily the mathematics of the instrument; it is the length of the horizon, the presence or absence of collateral, and whether the exposure is correlated with the counterparty's own creditworthiness.

FX spot and short-dated forwards sit at the bottom. A spot trade settles in two days and carries almost pure settlement risk, mitigated by CLS where the currency pair is eligible. A three-month forward carries a small replacement-cost exposure driven by the move in the forward rate over a short window. Volumes are enormous and tickets are numerous, so the operational challenge is throughput rather than modelling — the exposure per trade is small and the horizon is too short for the distribution to widen much.

Repo and securities financing transactions are next. A repo is a sale of a security with an agreement to buy it back, economically a secured loan. The exposure is the difference between the cash lent and the market value of the collateral held, which is small by construction and controlled through haircuts and daily margining. The difficulty here is not size but concentration and correlation. If the collateral is the debt of a name related to the counterparty, both legs fail together, which is specific wrong-way risk in its purest form. Liquidity is also the binding constraint: SFT books are enormous relative to their exposure, and a haircut that is right in normal markets can be badly wrong in a fire sale.

Exchange-traded and cleared derivatives replace the counterparty rather than the risk. When a futures contract or a standardised interest rate swap is cleared, a central counterparty interposes itself between the two original parties through novation, becoming buyer to every seller and seller to every buyer. Each participant now faces the CCP, which is highly rated, margins daily and initially, and stands behind a default waterfall funded by its members. The bilateral exposure largely disappears. What replaces it is exposure to the CCP itself, exposure to the clearing member if the bank clears through one, and — the more practically important item — a large and volatile demand for liquid margin. The risk has been transformed, not deleted.

Bilateral collateralised OTC derivatives under a Credit Support Annex are the industry's default state for financial counterparties. A CSA is the annex to the ISDA Master Agreement that governs collateral. It sets out which assets are eligible, what haircuts apply, the threshold below which no collateral is required, and the minimum transfer amount below which no call is made. Variation margin moves daily against the net mark-to-market of the whole netting set. The residual exposure is what can accumulate between the last good margin call and the moment the bank has finally closed out and replaced the trades — the margin period of risk, typically assumed at ten business days for uncleared portfolios. The difficulty is operational precision and dispute management rather than long-horizon modelling.

Bilateral uncollateralised long-dated OTC derivatives are the hardest rung, and they are the reference case for this module. Here there is no CSA and therefore no daily neutralisation. The exposure builds for years, driven by the diffusion of the underlying market factors, and the bank carries the whole distribution. Corporates are the typical counterparties, and often the ones with no traded credit default swap, which means the default probability has to be proxied and the resulting CVA cannot be hedged with a matching single-name instrument. Everything the bank has — simulation, limits, pricing, capital — is stressed simultaneously.

Structured and exotic derivatives with wrong-way features sit above even that. A derivative whose payoff depends on the credit or the equity of the counterparty itself, or on a market that drives the counterparty's solvency, produces an exposure that is largest exactly when recovery is lowest. Total return swaps on a concentrated equity portfolio, quanto and currency-linked structures written with counterparties in the same currency's sovereign risk, and credit derivatives referencing correlated names all belong here. The modelling requires a joint distribution of market and credit factors, the calibration is weakly supported by data, and the failures are not gradual.

Product Typical horizon Exposure profile Collateralised? Where the difficulty lies
FX spot and short forwards Days to months Small, short-lived, settlement-dominated Rarely; CLS mitigates settlement Throughput and intraday settlement limits
Repo and SFTs Overnight to 1 year Small, secured; gap risk on collateral Yes, daily, with haircuts Concentration, correlated collateral, fire-sale liquidity
Cleared and exchange-traded Days to 30 years Near zero bilaterally; margin-driven Yes, VM and IM to the CCP Liquidity demand and CCP/member concentration
Bilateral OTC with CSA 1 to 10 years Bounded by MPoR gap after margining Yes, daily VM, often IM Operational precision, disputes, MPoR assumptions
Bilateral uncollateralised OTC 5 to 30 years Large, diffusing, peaks mid-life No Long-horizon simulation, proxy credit, unhedgeable CVA
Structured and exotic, wrong-way 1 to 15 years Peaks when counterparty is weakest Varies, often partial Joint market-credit modelling, thin calibration data

What actually rises as you climb the ladder is the length of time between the decision and the consequence. At the bottom, a mistake in an FX spot book is discovered within days and costs a manageable amount. In the middle, a mispriced margin assumption on a collateralised book is discovered when a dispute escalates or when a close-out reveals that ten days was not enough. At the top, a wrong assumption about a proxy credit curve or about the correlation between a counterparty's solvency and the market driving its hedge can sit undetected for years and surface as a single large number. This is why seniority and independence are concentrated at the top of the ladder, and why the products that produce the least trade volume consume the most credit-committee time.

Two adjacent exposures belong in the same conversation even though they are not derivatives. A committed but undrawn credit line is a promise to lend, and the borrower will draw it precisely when its condition has deteriorated and other funding has closed — an option written by the bank, exercised against it at the worst moment, and measured through a credit conversion factor rather than a simulation. Settlement and delivery risk, discussed above, is the other: full-principal, short-duration and managed through payment-versus-payment infrastructure and daily settlement caps. Both are counterparty exposures in the plain sense that the bank stands to lose money because someone else did not perform, and both are usually held on different systems from the derivatives exposure, which is how a bank ends up with an incomplete view of what it is owed by a single client.

The product ladder maps almost exactly onto a systems ladder, and this is why counterparty-risk architecture is so fragmented. FX and settlement risk usually live in the payments and treasury platforms. SFTs live in a securities financing system with its own collateral logic. Cleared trades live wherever the clearing broker interface lives, with margin fed in from CCP files. Bilateral OTC lives in the derivatives trade capture system and flows to the exposure engine. Each was built by a different programme, in a different decade, against a different regulation. The aggregated group-level view of a counterparty is therefore almost never a single query — it is a reconciliation, and the hardest data problem in this module is making one counterparty in six systems resolve to one legal entity with one LEI.

The reference case is chosen to be the hardest one, not the typical one

A teaching example should sit where every control is under strain at once, and that is an uncollateralised long-dated cross-currency swap with a mid-investment-grade corporate. The running case for this module is Helvetia Chemicals AG, a Swiss specialty chemicals group with revenue of EUR 3.1bn and an internal rating equivalent to BBB−. It has no public credit default swap, so its default probability has to be proxied from the European chemicals BBB sector curve. It is a non-financial counterparty below the EMIR clearing threshold, an NFC−, which means it is exempt from the clearing obligation and from the uncleared margin requirements. That exemption is the whole reason the case teaches: the regulatory architecture built after 2008 was designed to collateralise and centrally clear derivatives risk, and the corporate segment was deliberately left outside it so that companies could keep hedging without tying up cash in margin accounts. The risk did not vanish. It stayed on bank balance sheets, uncollateralised, and it is where CVA still lives.

Four features stack on top of each other to make this the hardest case in the ladder. The tenor is ten years, so the market factors have a decade to diffuse and the exposure distribution is wide. There is no collateral, so nothing truncates that distribution — the bank carries the full tail. There is an exchange of notional at maturity, which means the largest single cash flow in the trade is also the last one, and the exposure profile does not decay towards zero the way an interest rate swap's does; it stays elevated deep into the life of the trade. And two market factors drive it simultaneously — the EUR/USD spot rate and the interest rate differential between the two currencies — with FX volatility dominating at long horizons.

On top of that sits a credit problem with no clean solution. Helvetia has no traded CDS. The default probability that goes into the CVA calculation is inferred from a sector proxy curve, which means the single largest input to the charge is an estimate the bank cannot observe or verify in the market. It also means the resulting CVA cannot be hedged with a matching single-name credit default swap. The XVA desk can hedge the market sensitivity of the exposure profile and can take an index CDS position against broad credit spread moves, but the idiosyncratic component of Helvetia's credit — precisely the part that will move if Helvetia is downgraded — sits unhedged on the bank's books for a decade.

The exemption that makes Helvetia hard is a policy choice with a defensible rationale, and it is worth understanding rather than dismissing. When EMIR was drafted, the argument from the corporate sector was straightforward. A manufacturer hedging a genuine commercial exposure — a currency mismatch between debt and revenue, a fuel cost, an interest rate on a capital project — is reducing risk, not taking it. Forcing that manufacturer to post daily variation margin would require it to hold a liquidity buffer against the mark-to-market of its own hedge, which would raise the cost of hedging and, at the margin, cause some companies to stop hedging altogether. That would leave the real economy more exposed, not less. So non-financial counterparties whose speculative derivative positions fall below defined clearing thresholds were exempted from the clearing obligation and from the margin requirements.

The rationale is sound and the consequence is that the risk did not disappear — it concentrated. Every uncollateralised corporate hedge is a long-dated unsecured credit exposure sitting on a bank balance sheet, priced through CVA, capitalised under SA-CCR or IMM, and funded through FVA. The corporate segment is therefore where counterparty risk is at its most economically significant per unit of notional, and where the machinery described in this module does the most work. Cleared and margined portfolios generate enormous volumes and comparatively little exposure. A book of NFC− hedges generates modest volumes and most of the CVA. That inversion is the reason a module on counterparty risk built around cleared products would teach almost nothing.

Part 2 walks this trade through all ten stages of the lifecycle. The exposure numbers, the CVA charge and its decomposition, the collateral cascade, the capital calculation and the close-out scenario all belong there and are not previewed here. What Part 1 establishes is only why this particular trade is worth ten stages of attention: it is the case where an exemption in the rules, a long horizon, an absent hedge and an unobservable input meet in a single ticket.

Notional is not exposure, and netting is the largest risk reducer in the system

The headline number for the derivatives market is enormous and almost entirely misleading. The Bank for International Settlements publishes semi-annual statistics on over-the-counter derivatives, and the notional amount outstanding of global OTC derivatives has run in the region of USD 600 to 700 trillion in recent years. That figure is quoted constantly and understood rarely. Notional is the reference amount used to calculate payments; in an interest rate swap it is never exchanged at all. It measures the scale of the contractual plumbing, not the amount of money at risk.

Gross market value is a small fraction of notional, and gross credit exposure is smaller again. Gross market value — the sum of the absolute replacement values of all outstanding contracts, before any offsetting — has run in the single-digit trillions of dollars, a percentage point or two of notional. Gross credit exposure, which is gross market value after legally enforceable bilateral netting, is smaller again, typically well under half of gross market value. The three numbers describe the same market and differ by roughly two orders of magnitude from end to end.

Measure What it actually is Approximate global scale What it tells you
Notional outstanding Reference amount for calculating payments USD 600–700 trillion (BIS) Size of the contractual book, not of risk
Gross market value Sum of absolute replacement values, no offsetting Single-digit USD trillions What the contracts are worth today
Gross credit exposure Gross market value after enforceable netting A fraction of gross market value What is genuinely at risk before collateral
After collateral Gross credit exposure less VM and IM held Materially smaller again What the bank is left carrying

The gap between the second and third rows is the single most consequential number in this module. It is produced by one legal mechanism — close-out netting — and by nothing else. Under an ISDA Master Agreement, all transactions between two parties are deemed to form a single agreement. If one party defaults, the non-defaulting party terminates every transaction, values each one, and adds them up with sign, arriving at a single net amount owed in one direction. Without that clause, a bankruptcy administrator could do the opposite: demand payment on every contract where the estate is in the money and leave the bank to claim as an unsecured creditor on all the ones where it is. That practice has a name in the industry, cherry-picking, and it is exactly what close-out netting exists to prevent.

Netting converts a portfolio of gross claims into one net claim, and the arithmetic reduction is brutal. A counterparty with four hundred trades, some deeply in the money and some deeply out, may show tens of millions of gross positive value and a small net figure once the negatives are set against the positives. Collateral then attacks what is left. The order matters and it is always the same — net first, then collateralise the remainder — which is why the netting set, and not the individual trade, is the unit of risk in this business. It is also why the legal work described in Stage 2 of the lifecycle is not a compliance formality. A netting opinion that fails in a given jurisdiction does not increase the exposure by a few percent; it can multiply it several times over, instantly, on a portfolio the bank thought it understood.

The arithmetic is worth doing once slowly, because the intuition it builds is the one Part 2 pays off. Take a counterparty with which the bank has a mixed book — some swaps in the money, some out. Suppose the positive marks add to EUR 142m and the negative marks to EUR 104m. Without enforceable netting, the exposure is EUR 142m, because the administrator would collect on the EUR 104m the estate is owed and leave the bank claiming as an unsecured creditor on the EUR 142m. With enforceable netting, the exposure is EUR 38m. Variation margin already held then reduces it further, and initial margin held against the residual gap can bring it close to zero. The reduction from EUR 142m to something near nothing is a cascade of four steps, and the largest single step is the first one — the legal one. Those figures are illustrative and are used again in Part 2's treatment of collateral.

Two conditions have to hold for any of that to be true, and neither is automatic. The first is that the trades must be inside the same netting set, meaning they are governed by the same master agreement between the same two legal entities. Trades with a parent do not net against trades with a subsidiary. Trades under a second master agreement, signed by a different desk years earlier and never consolidated, do not net against the first. The second condition is that the netting must be enforceable in the jurisdictions that matter — the counterparty's place of incorporation, the governing law of the agreement, and any jurisdiction where insolvency proceedings could plausibly be opened. Banks maintain, and pay external counsel to refresh, netting opinions covering dozens of jurisdictions and entity types, and where an opinion is absent or qualified the bank must calculate exposure gross. A single unfavourable opinion on a large book moves the exposure number by a multiple, not a margin.

CVA is an accounting number, which is why it hurts

IFRS 13 requires fair value to reflect the credit risk of the parties to a contract, and that single requirement moved CVA out of risk management and into the income statement. Before it, a bank could hold a prudent reserve against counterparty losses and treat changes in it as an internal matter. After it, the fair value of a derivative portfolio must include the market's assessment of counterparty credit, so when a client's credit spread widens the bank's derivative assets are worth less and the difference is recognised in reported earnings. No default has to occur. Nobody has to miss a payment. A repricing of credit risk in the market is enough to produce a loss that shareholders see.

That is why banks built CVA desks rather than CVA reserves. A number that hits reported earnings and moves with observable market factors is, by construction, a trading position, and it is managed by traders. A large dealer's XVA desk is typically a team of tens rather than hundreds — traders, quantitative analysts and dedicated technology staff — sitting on top of a computational platform that is among the most expensive the bank runs. The desk is small relative to the size of the exposure it manages, and its influence comes from the price it quotes internally rather than from headcount.

CVA also has a mirror image, and it is one of the stranger objects in bank accounting. If the bank must value the risk that its counterparty fails, symmetry says the counterparty must value the risk that the bank fails — and under fair-value accounting the bank recognises that too, as a debit valuation adjustment. DVA is a gain to the bank when its own credit spread widens, because the contracts it owes are less likely to be paid in full. The economics are uncomfortable and the hedging is worse, since a bank cannot sell protection on itself. Regulators responded by filtering DVA gains out of regulatory capital even though accounting standards leave them in reported earnings, which is one of several places in this module where the accounting answer and the capital answer diverge on purpose. Part 2 sets out the full XVA family and what each member compensates for.

The technology behind all of this is a cost centre justified by capital relief, not by revenue. An exposure engine does not generate a spread. The business case for rebuilding one is made in the language of regulatory capital — a more accurate netting-set treatment reduces exposure at default, an approved internal model reduces the multiplier applied to it, a better collateral system reduces the residual exposure that attracts a charge — and in the language of avoided losses and avoided supervisory findings. This has a lasting consequence for how these platforms are funded and therefore how they are built. They are approved when a deadline forces them, sized to the deadline, and rarely refactored afterwards.

The regulatory cycle, not the market cycle, drives the build cycle

Figure 1 Counterparty-risk investment follows the rulebook, not the market cycle. Vertical scale is illustrative.

Most businesses inside a bank invest when revenue justifies it; counterparty risk invests when a deadline requires it. The consequence is that every platform in this module carries visible sediment from each regulatory wave it has survived, and the timeline of those waves explains more about how a bank's systems look than any architectural principle does.

2008 is the origin, and two failures define it. Lehman Brothers' bankruptcy in September 2008 forced the close-out of roughly 900,000 derivative transactions across some 6,500 legal entities, and the resulting disputes over close-out valuations ran through the courts for more than a decade. It demonstrated that the mechanics of terminating a large derivatives book had never been tested at scale, and that close-out netting worked cleanly where documentation was clean and failed where it was not.

AIG failed differently, and its mechanism is the one worth understanding precisely. AIG's Financial Products unit had sold enormous volumes of credit protection through credit default swaps on structured credit, largely uncollateralised, on the strength of the group's AAA rating. The contracts contained ratings-linked collateral triggers. As the underlying mortgage collateral deteriorated, the mark-to-market of those swaps moved sharply against AIG, and its counterparties called for collateral. When AIG was downgraded, the triggers required it to post substantially more. It did not have the liquidity to meet the calls, and the resulting funding gap — not an actual default on the protection it had written — is what brought it to the point of requiring government support. AIG is therefore the founding case for the proposition that collateral converts credit risk into liquidity risk. The obligation to post is unconditional and immediate; the ability to post is not.

2010 brought the Basel III CVA capital charge, and it rested on one statistic. The Basel Committee found that roughly two-thirds of counterparty-credit-related losses during the 2007–09 crisis came from mark-to-market moves in CVA rather than from actual defaults. Banks had been holding capital against the event of a counterparty failing, and had been losing money instead on the repricing of counterparties that survived. The new charge required capital against CVA volatility itself. That was the moment counterparty risk stopped being a credit discipline and became a market-risk discipline with a credit input, and it is directly why XVA desks exist in their current form.

2013 brought the EMIR clearing obligation in the EU, alongside Dodd-Frank Title VII in the US. Standardised OTC derivatives — principally interest rate swaps and index credit default swaps — had to be cleared through a central counterparty. Trade reporting became mandatory. Risk mitigation techniques, including timely confirmation, portfolio reconciliation and dispute resolution, became legal requirements for trades that remained bilateral. The build implications were large and lasting: clearing connectivity, CCP margin processing, a reporting pipeline to trade repositories, and the beginnings of a serious reference-data problem as every counterparty needed a Legal Entity Identifier.

From 2016 to 2022 the Uncleared Margin Rules phased in. The BCBS-IOSCO framework required variation margin for all in-scope counterparties and, more disruptively, initial margin — an amount posted upfront against potential future exposure, segregated with a third-party custodian and not available for rehypothecation. It arrived in phases by size, starting with the largest dealers and ending in September 2022 with counterparties whose aggregate average notional amount exceeded EUR 8bn. ISDA SIMM became the standard model for calculating it. Each phase pulled a new tier of institutions into a build they had not previously needed, and initial margin created an entirely new funding cost, MVA, which had to be modelled and priced.

2021 brought SA-CCR into force in the EU under CRR2, and it changed the arithmetic of every derivative desk. The Standardised Approach for Counterparty Credit Risk replaced the Current Exposure Method, a formula that had barely changed since 1988 and that recognised netting and collateral only crudely. SA-CCR is more risk-sensitive: it distinguishes margined from unmargined trades, recognises collateral properly, and separates exposure into a replacement cost and an add-on for future movement. It is also, for many portfolios, considerably more expensive, and it made long-dated uncollateralised trades — the Helvetia case — visibly costly in capital for the first time. Part 2 works the formula through in full.

2022 delivered the crisis nobody had modelled, and it was a liquidity crisis inside a credit framework. European energy firms faced tens of billions of euros of margin calls as gas and power prices spiked, and several governments extended emergency liquidity facilities to companies that were solvent but could not fund the collateral their hedges demanded. In the United Kingdom, the September gilt sell-off forced liability-driven investment funds serving pension schemes into selling the very assets that were falling in order to meet collateral calls, and the Bank of England intervened. Both episodes made the same point from opposite directions: a decade of policy had successfully converted credit risk into collateral, and had thereby created a system in which the binding constraint is the ability to post, not the ability to pay. Part 5 develops both cases.

FRTB-CVA and Basel 3.1 close the sequence for now. The revised CVA risk framework replaces the original Basel III methods with a basic approach and a sensitivity-based standardised approach, demanding sensitivity calculations at a granularity that many legacy engines were never designed to produce. Basel 3.1 introduces an output floor limiting how far a bank's internal models may reduce capital below the standardised result, which narrows the payoff from an approved internal model and quietly changes the buy-versus-build calculation described in Part 4.

Read the timeline as an architecture diagram and it explains most of what you will find inside a bank. Each wave arrived with a deadline, and a deadline funds a project rather than a platform. The result is a stack in which a 2013 reporting pipeline sits beside a 2016 margin engine and a 2021 capital calculator, each built to satisfy a different rule, each holding its own copy of the same trade, and each reconciling to the others through processes that exist only because the systems were never designed together. That is not incompetence. It is the predictable shape of a business whose investment cycle is set by a rulebook.

The regulatory sediment has a specific technical signature, and it is worth learning to spot. Look for the number of distinct trade repositories inside one bank: a front-office trade store, a risk-engine trade snapshot, a collateral system position file, a capital calculator input, and a regulatory reporting extract. Each was built for one rule. Each is authoritative for its own purpose and wrong for every other. The reconciliation layer between them is usually the largest single line item in a counterparty-risk technology budget, and it produces nothing a regulator asked for. Removing it is the highest-value modernisation available in this domain, and it is almost never funded on its own, because no deadline requires it.

  1. Why is the exposure on a ten-year swap harder to manage than the exposure on a ten-year loan of the same size?

  2. What is settlement risk, and why is it treated as a separate category?

  3. Why does the XVA desk sit in the front office rather than in risk management?

  4. A client asks why the bank charges it more for a swap than it charges a leveraged hedge fund with a worse rating. What is the answer?

  5. Roughly two-thirds of counterparty-credit losses in the 2007–09 crisis came from CVA mark-to-market moves rather than from defaults. Why does that statistic matter so much?

  6. Why was AIG's failure a liquidity failure rather than a credit failure?

  7. Why are counterparty-risk platforms so often a patchwork?

1

  1. Why is a ten-year swap's exposure harder to manage than a ten-year loan's?

  2. Why does the XVA desk sit in the front office?

  3. Why is settlement risk a separate category?

2The lifecycle: from a client who wants a swap to a bank that has to be paid

A derivative is agreed in an afternoon and lived with for a decade. This part walks the whole decade, stage by stage, and shows where the money is actually won and lost.

The ten stages are one process, not ten departments

A counterparty-risk lifecycle is not a workflow diagram; it is a chain in which every link inherits the mistakes of the link before it. The legal entity chosen at onboarding determines which netting opinion applies. The netting opinion determines whether the exposure can be netted. The netted exposure determines the CVA. The CVA determines the price quoted to the client. The price quoted to the client determines whether the bank makes money over ten years. Get the entity wrong in week one and the error surfaces as a valuation dispute in an insolvency court in year eight.

The ten stages below are the spine of this module. Every other part refers back to them by these names and numbers. Read the table once now; you will meet each row again in detail.

Stage Duration Core question answered Output Who leads
1. Onboarding and credit assessment 2–12 weeks Who is this counterparty and how likely is it to default? Internal rating, PD, approved relationship Credit risk
2. Legal architecture — ISDA, CSA, netting opinions 4–16 weeks If they default, what am I legally entitled to claim? Signed ISDA, Schedule, CSA, clean netting opinion Legal and documentation
3. The limit framework Set annually, reviewed continuously How much of this counterparty am I willing to own? Limit grid by entity, product, tenor Credit risk and credit committee
4. The pre-deal check Seconds Does this specific trade fit inside the limits? Pass, fail, or escalate Front office and the limit engine
5. XVA pricing and the charge Minutes What does this risk cost, and who pays for it? The XVA charge embedded in the client price XVA desk
6. Booking and trade capture T+0 Is the trade in the systems, correctly, once? Confirmed trade record with correct static data Operations
7. The daily exposure cycle Every night What am I owed, and what could I be owed? EE, EPE, PFE profiles per netting set Risk engine and market risk IT
8. Collateral and margin Daily, intraday in stress How much of the exposure can I convert into cash today? Margin calls made, received, disputed, settled Collateral operations
9. XVA hedging and capital Continuous How do I stop the CVA moving my P&L, and what capital does this consume? Hedges executed, EAD and RWA reported XVA desk and capital management
10. Distress, default and close-out Days to years What do I actually recover? Close-out amount, net claim, realised loss Credit, legal and workout

The pressure is concentrated in two moments, and everything between them exists to make the second one survivable

An issuance process has a deal day; counterparty risk does not. In a bond issuance or an IPO there is a single afternoon when the price is struck, the book is allocated and the outcome is known within days. Counterparty risk has no such day. Instead it has two moments of concentrated consequence, ten years apart, and a long grinding middle whose entire purpose is to make the second moment survivable.

The first moment is Stage 5, the pricing decision. When the bank quotes a cross-currency swap it fixes the credit charge for the whole life of the trade. There is no annual repricing, no margin reset, no covenant that lets the bank raise the charge when the client deteriorates. If the CVA is under-priced by twenty basis points, that error is locked in for a decade and cannot be recovered from the client. It can only be absorbed by the bank or hedged in the market at a cost the bank has already declined to charge for. This is the sense in which counterparty risk is different from lending: a loan can be repriced at rollover, a ten-year swap cannot.

The second moment is Stage 10, the close-out. This is where the loss stops being a reserve and becomes cash. Everything the bank did in between — the daily simulation, the margin calls, the limit reviews, the hedges — was an attempt to make the number at close-out smaller than it would otherwise have been. None of that activity generates revenue on its own. It is pure loss mitigation, which is why it is chronically under-invested in until a bank has had a bad close-out.

Between the two moments sits nine and a half years of daily process that no client ever sees and no salesperson ever thanks anyone for. The nightly batch, the margin calls, the disputes, the limit utilisation reports, the annual credit reviews. Each individual day is low-consequence. The cumulative effect is enormous, because the quality of that daily process is what determines whether the bank arrives at the close-out with clean documentation, current collateral and an accurate view of what it is owed — or arrives with a stale netting opinion, three days of unsettled margin and a valuation it cannot defend.

Figure 2 The counterparty-risk lifecycle: two moments of consequence separated by a decade of daily process.

Step 1 — Onboarding decides what you are trading with, and it is almost never what the salesperson thinks

Before a bank can measure how much it might lose, it has to know precisely who owes it the money. That sounds trivial and it is not. A large corporate group is a tree of legal entities: a listed holding company, a financing subsidiary in the Netherlands, an operating company in Switzerland, a treasury vehicle in Ireland. Each is a separate legal person with its own balance sheet, its own creditors and its own insolvency estate. The salesperson says "we trade with Helvetia". The credit officer has to know whether the counterparty is Helvetia Chemicals AG, the Swiss parent, or Helvetia Finance BV, a Dutch subsidiary with EUR 200,000 of share capital and a parent guarantee that may or may not be enforceable.

This is the legal-entity hierarchy problem, and it is the first place counterparty risk goes wrong. The economic risk sits with the group — that is where the cash flows and the assets are. The legal claim sits with whichever entity signed the ISDA. If the bank faces a thinly capitalised subsidiary without a guarantee, it has group-level risk in its rating model and subsidiary-level recovery in an insolvency. Banks address this with a formal legal-entity hierarchy: every trading counterparty is mapped to an ultimate parent, exposures are aggregated at group level for limit purposes, and any reliance on a guarantee is documented and legally opined in the same way as netting.

The Legal Entity Identifier (LEI) is the market's answer to the identification half of this problem. It is a twenty-character alphanumeric code, issued under a global governance framework, that uniquely identifies a legal entity in financial transactions. It carries reference data on the entity's direct and ultimate parent. Regulators mandate it for trade reporting under EMIR and Dodd-Frank precisely because, before it existed, no supervisor could reliably aggregate one group's derivative exposures across venues and jurisdictions. Treat the LEI as a risk control, not a reporting field. An entity onboarded under the wrong LEI is an exposure that will never aggregate correctly, and the aggregation error is invisible until it matters.

Know Your Customer is a risk control disguised as a compliance task

KYC establishes who owns and controls the counterparty, where its money comes from, and whether it or its beneficial owners are sanctioned or politically exposed. In most banks it is run by a separate financial-crime function on a separate timetable, and front office treats it as an obstacle. That framing is wrong in a specific way that matters for counterparty risk: the same investigation that establishes beneficial ownership also establishes the group structure, and the group structure is the basis of aggregation. A KYC file that stops at the immediate counterparty leaves the credit function guessing at the tree.

The internal rating is an estimate of probability, expressed on a scale

A bank does not rate a counterparty in order to describe it; it rates a counterparty in order to attach a number to it. That number is the probability of default (PD): the probability, expressed as a percentage, that the counterparty fails to meet its obligations over a defined horizon, conventionally one year. A PD of 0.30% means that out of a thousand counterparties of this quality, roughly three are expected to default within the year. PD is a property of the obligor, not of any particular transaction.

The second number is loss given default (LGD): the fraction of the amount owed that the bank expects to lose if default occurs, after recoveries, expressed as a percentage. LGD is one minus the recovery rate. If a defaulted claim ultimately pays forty cents on the euro, the recovery rate is 40% and the LGD is 60%. LGD depends on the claim, not the obligor — a secured claim and an unsecured claim on the same defaulted company have very different LGDs.

The distinction is worth stating precisely because it is the source of persistent confusion. PD answers "will they fail?" LGD answers "if they fail, how much of my money disappears?" The obligor rating captures the first. The facility view — seniority, security, guarantees, collateral — captures the second. For most uncollateralised OTC derivatives the bank ranks as a senior unsecured creditor, which is why 40% recovery and 60% LGD are the conventional working assumptions, and why they are also the assumptions embedded in standard CDS contract conventions.

Internal rating scales typically run to fifteen to twenty-five grades, each associated with a PD range. A master scale maps those internal grades to the familiar external agency notation so that a bank's grade 9 can be described as "BBB− equivalent" in a committee paper. The mapping is a communication device, not an equivalence: the bank's own default history, not the agency's, calibrates the PDs. Investment-grade ratings for large corporates typically carry one-year PDs between roughly 0.02% and 0.35%; sub-investment-grade PDs rise steeply from there.

Financial spreading and the credit approval memo

Financial spreading is the process of taking a counterparty's published accounts and re-casting them into the bank's own standardised template, so that leverage, interest cover, cash conversion and liquidity can be compared across counterparties on a consistent basis. It is mechanical work, increasingly automated, and it is the raw material of the rating model. The output feeds a quantitative score, which the analyst then overrides — up or down, with written justification — on qualitative grounds: management quality, sector cyclicality, the group's willingness to support a subsidiary, refinancing walls.

The result is a credit approval memo: a document setting out the counterparty, the proposed limits, the rating and its rationale, the product scope, the collateral and documentation terms, and the credit officer's recommendation. For a new relationship in derivatives, it goes to a new-counterparty committee — a standing forum combining credit, legal, compliance and the business — which approves the relationship as a whole, not any individual trade. The committee's real function is to force a decision about appetite before commercial momentum builds. Once a salesperson has told a client "we can do this", declining becomes a relationship problem rather than a risk decision.

Onboarding is a data-creation event, and the data it creates is load-bearing for a decade. The counterparty master record — LEI, legal entity hierarchy, country of incorporation, jurisdiction of the governing law, industry classification, internal rating, group parent — is written once here and consumed by the limit engine, the exposure engine, the collateral system, the capital calculator and every regulatory report the bank files.

Two architectural points decide whether this works. First, there must be a single golden source for counterparty reference data, with every downstream system consuming it rather than maintaining its own copy; banks that let the trading system, the risk system and the collateral system each hold their own entity table spend the next decade reconciling them. Second, hierarchy must be modelled as a versioned graph with effective dates, not a flat parent field, because corporate structures change and last year's exposure aggregation has to remain reproducible.

Typical stack: a vendor or in-house client master, an LEI reference feed from a registration authority, a KYC case-management workflow, a financial-spreading tool feeding the rating model, and an entitlement layer that stops anyone booking a trade against an entity that is not fully onboarded.

Step 3 — Limits are the bank's statement of appetite, and they are set on the tail rather than the average

A limit is the only place in a bank where risk appetite becomes a number that can stop a trade. Everything else — policy, culture, committee minutes — is advisory. The limit is binding, checked mechanically, and enforced before execution. That is why the limit framework is the internal governance gate of this whole business, and why the two most instructive counterparty-risk failures of the last two decades were both, at root, limit-framework failures rather than modelling failures.

The limit hierarchy is a set of nested constraints, each answering a different question

A single number cannot express appetite for a counterparty, because the bank is worried about several different things at once. It is worried about the counterparty failing. It is also worried about being unable to exit a long-dated position, about being over-exposed to a sector when the sector turns, and about a settlement failing on a single day for an amount far larger than the average exposure. Each concern gets its own limit.

Limit type What it constrains The question it answers
Group limit Total exposure across all entities of a corporate group How much of this credit do I own in total?
Entity limit Exposure to one legal entity How much can I actually claim in that entity's insolvency?
Product limit Exposure by product family — FX, rates, equity, commodity Am I taking risk I can measure and hedge?
Tenor bucket limit Exposure beyond stated maturities, for example over five years Can I exit if I need to, and how long am I locked in?
Settlement limit Gross amount settling on any single day Can I survive a delivery failure today?

The tenor bucket limit deserves particular attention because it is the one most often waived. A ten-year uncollateralised swap is not a five-year swap with more time on it. It carries a rating migration risk that compounds: a BBB− counterparty has a roughly 0.3% one-year PD but a cumulative ten-year default probability approaching 20%, and over that horizon most counterparties that default were investment grade when the trade was struck. It also has no exit. There is no liquid market in which a bank can sell a bilateral swap with a mid-sized corporate. Capping long-dated exposure separately is the bank's admission that time itself is a risk factor.

Settlement limits sit slightly apart from the rest because they measure a different thing entirely — the gross principal changing hands on a single value date, not the mark-to-market. For the Helvetia swap, the ten-year exposure profile peaks at EUR 68m, but on 15 March 2036 the bank delivers EUR 460m and expects USD 500m back. If Helvetia fails between the two legs settling, the bank's loss is the full principal, not the mark-to-market. That is Herstatt risk, and it is why settlement limits are measured in hundreds of millions where PFE limits are measured in tens.

Figure 3 A limit is a nest of containers, and a trade must clear every one of them. Illustrative figures.

A limit is not one number, and a trade does not pass one test. Helvetia sits inside four constraints at once and the swap has to clear all of them. The group carries a PFE limit of EUR 250m, of which EUR 168m is already used across three subsidiaries, a comfortable 67%. The Swiss operating entity that actually faces the bank on this trade carries EUR 75m, and after the swap it stands at EUR 68m, or 91%. The tenor bucket beyond five years carries EUR 45m and runs at EUR 41m, also 91%. The daily settlement limit of EUR 120m is barely touched at EUR 22m, because settlement risk on a swap is a coupon exchange rather than a principal transfer. Every figure here is illustrative.

The binding constraint is almost never the one people quote. The group headline of 67% is the number that travels upward in reporting and it is the least informative of the four. This trade is constrained by the entity limit and by the long-tenor bucket, both at 91% and both above the bank's 90% escalation trigger, which means it does not get done by a salesperson and a credit officer alone. A framework that reports only the outermost container will be surprised by its own book.

Limits are set on PFE, not on notional and not on current mark-to-market

Notional is not exposure. A EUR 460m cross-currency swap does not put EUR 460m at risk; it puts at risk the difference in value between two streams of payments, which is a fraction of the notional. Limiting on notional would treat a ten-year swap and a one-week forward of the same size as identical, which is absurd.

Current mark-to-market is not exposure either, and this is the subtler error. At inception a swap is priced at fair value, so its mark-to-market is zero. A limit framework based on current MtM would show zero utilisation for a trade the bank will still hold in 2036. It measures what happened, not what could happen.

The measure that works is potential future exposure (PFE): the largest amount the bank could plausibly be owed at a future date, at a chosen confidence level. PFE is a quantile of the simulated distribution of future exposure. A 97.5% peak PFE of EUR 68m means that, across the simulated futures, in only 2.5% of them does the exposure at the worst point of the profile exceed EUR 68m.

The choice of percentile is a policy decision, not a technical one, and banks should be honest about that. At 95% the bank is defending against a one-in-twenty outcome; at 99% against a one-in-a-hundred. Moving the confidence level from 95% to 99% on a long-dated FX position can raise the reported PFE by fifty per cent or more, with no change whatsoever in the underlying risk. Nothing about the trade has changed. What has changed is how far into the tail the institution has decided to defend. A bank that sets limits at 95% and a bank that sets them at 99% are making different statements about appetite, and the numbers in their limit reports are not comparable.

Concentration limits address the risk that individual limits are all fine

A portfolio of individually acceptable exposures can be an unacceptable portfolio. Twenty European chemicals counterparties each within limit is a single bet on European chemicals. Concentration limits address this on three conventional axes — single name, sector and country — by capping the aggregate exposure to any one of them as a proportion of capital or of the total book.

The Archegos episode in March 2021 exposed a fourth axis that most frameworks did not have. Archegos Capital Management was a family office that built very large concentrated equity positions through total return swaps with several prime brokers at once. Each prime broker saw its own exposure, which was within its own limits. None saw the aggregate. When the underlying stocks fell and the positions could not be liquidated in an orderly way, Credit Suisse lost roughly USD 5.5bn, Nomura roughly USD 2.9bn and Morgan Stanley roughly USD 900m.

The mechanism is worth stating carefully because it generalises. Where a client trades the same economic position with several dealers, each dealer's exposure is a function of a market factor — the price of the underlying — that every other dealer's exposure is also a function of. The exposures are perfectly correlated, and so is the liquidation. When all the dealers try to exit at once, the market impact each of them assumed in their own model applies to all of them simultaneously. The invisible aggregate is not just a measurement gap; it is a liquidity assumption that is only valid if you are the only one making it. Part 5 develops the case. For limit-setting purposes the lesson is that concentration must be assessed against the client's plausible total position, not the bank's share of it, and that a client who refuses to disclose its positions elsewhere is telling you something.

The delegated authority grid decides who can say yes

Not every limit decision goes to a committee. Banks operate a delegated authority grid that scales the approval level to the size, tenor and rating of the exposure.

Approver Typical remit Turnaround
Desk head Temporary excesses within pre-agreed tolerance on approved names Same day
Credit officer Limits for rated counterparties within standard product and tenor bands Days
Credit committee New relationships, sub-investment grade, long tenor, non-standard structures Weekly cycle
Board risk committee Framework, appetite statement, largest single exposures, policy exceptions Quarterly

The grid matters more than its contents. What makes a limit framework credible is that the person who benefits from the trade is never the person who approves the limit for it. Where that separation erodes — where a business head can approve their own excesses, or where a credit function reports into the business — the framework becomes documentation of decisions rather than a control on them.

Limits are reviewed at least annually, and continuously in substance. A watchlist captures counterparties whose credit quality has deteriorated but which have not defaulted: rating downgrades, spread widening, covenant pressure, auditor concerns, adverse news. Watchlist status typically freezes new business, tightens limits, and triggers a plan for reducing the existing position. Step 10 picks this up.

The limit system is a real-time transactional system pretending to be a reporting system, and this is the source of most of its problems. It has to hold the limit hierarchy, consume exposure numbers from the overnight risk engine, apply intraday updates from trade capture, respond to pre-deal enquiries in well under a second, and maintain a defensible audit trail of every check, breach, approval and override.

The hard part is the intraday reconciliation between the overnight PFE profile and the day's new trades. The clean design keeps the overnight profile as the baseline and applies incremental adjustments as trades book, then reconciles to the next overnight run and investigates the drift. The failure mode is a limit system whose numbers are twenty-four hours old and whose users know it, because they then stop treating a breach as real.

Concentration reporting needs the entity hierarchy from Stage 1 and a consistent sector and country taxonomy, and it needs them to be the same taxonomy the capital calculation uses. Where they differ, the bank's concentration report and its regulatory report will not agree, and neither will be trusted.

  1. Why is a limit set on potential future exposure rather than on current mark-to-market?

  2. Helvetia's group limit is 67% used and its entity limit is 91% used. Which one governs?

  3. What does a concentration limit catch that a counterparty limit cannot?

Step 4 — The pre-deal check answers a question about the portfolio, not about the trade

When a salesperson asks whether a trade can be done, the credit system must not answer by computing how much exposure that trade creates. That is the wrong question, and answering it produces systematically wrong decisions. The right question is what this trade does to the exposure profile of the netting set it will join.

The distinction is between standalone exposure — what the trade's exposure profile looks like on its own, as if it were the only trade with that counterparty — and incremental exposure, the difference between the netting set's exposure profile with the trade and without it. For an empty netting set the two are identical. For a netting set with existing positions they can be very different, and the incremental number can be negative.

A concrete case. The bank already holds a five-year EUR/USD cross-currency swap with a client in which it receives USD. The client now asks for a five-year swap in the opposite direction. Standalone, the new trade has a substantial exposure profile. Incrementally, it largely cancels the existing one: in the states of the world where the old trade is deep in the money to the bank, the new trade is deep out of the money, and the net portfolio value is close to flat. The incremental PFE is negative. Booking the trade reduces the bank's exposure and should reduce its limit utilisation and its CVA.

This is not an edge case; it is the normal condition of an active netting set, and it is why exposure must be computed at netting-set level and allocated back to trades rather than computed at trade level and summed. A bank that measures standalone exposure will decline risk-reducing trades and will over-charge clients who are hedging positions they already have with it — which is to say, it will systematically drive away its best-behaved clients.

Latency is the binding constraint, and it shapes the architecture

A full Monte Carlo exposure calculation for one netting set — thousands of paths, dozens of time steps, every trade repriced at every node — takes minutes to hours. A salesperson on the phone with a client has perhaps two seconds. This gap is the central engineering problem of Stage 4, and banks solve it in three ways, usually all three at once.

Pre-computed grids. The overnight batch stores not just the exposure profile but the netting set's value at every node of the simulation. A new trade's value can then be computed at those same nodes and added, giving an approximate new profile without re-simulating the market. This is by far the most common design and it is accurate when the new trade is priceable quickly.

Add-on approximations. For standard products the system holds a table of standalone PFE add-ons by product, tenor, currency and notional, adjusted by a netting-benefit factor estimated from the existing portfolio. Crude, fast, and adequate for small trades against large books.

Cut-down simulation. A reduced path count and a coarse time grid produce a real answer in seconds rather than minutes, with wider error bars. Increasingly, banks run this on demand using GPU acceleration and treat the overnight run as the reconciliation baseline rather than the source of truth.

The trade-off is explicit and should be governed explicitly. A faster answer is a less accurate answer, and the tolerance for inaccuracy should scale inversely with how close the trade takes the counterparty to its limit. A trade that takes utilisation from 20% to 24% does not need precision. A trade that takes it from 88% to 91% does.

The breach workflow decides whether the framework is real

When the check fails, one of four things happens, and which one determines whether limits mean anything in that institution. The trade is declined. The trade is restructured — shortened, reduced, collateralised, or given a break clause — until it fits. A temporary excess is approved within delegated authority, with an expiry date and a documented reason. Or the request escalates to the credit officer or committee, and the client waits.

The pathology to watch for is the permanent temporary excess. An excess granted for a week, rolled for a month, rolled for a year, is a limit increase that never went to the body that should have approved it. Banks control this by making excesses expire automatically rather than requiring cancellation, by reporting the age distribution of open excesses to the risk committee, and by requiring escalation to a higher authority each time an excess is rolled.

The pre-deal check is the only piece of the counterparty-risk stack with a hard latency budget, and it should be engineered like a pricing service rather than like a risk report. Target response is a few hundred milliseconds, which rules out reading a relational database at request time and rules in an in-memory cache of netting-set profiles, refreshed from the overnight batch and updated incrementally as trades book.

Three requirements are non-negotiable. Every check must be logged with its inputs, its result and its approver, because this is the audit trail the regulator asks for. The service must degrade safely — if the exposure cache is stale or unavailable, it should escalate rather than approve. And it must be the same service that the electronic trading channels call, so that a trade priced by an algorithm faces the same gate as a trade priced by a person.

Step 5 — The CVA charge is a price, not a fee

This is the moment of irreversible consequence. The bank is about to fix, for ten years, what it charges for the possibility that Helvetia fails. Nothing later in the lifecycle can revise it. The entire apparatus of Stages 1 to 4 exists to make this number computable, and the entire apparatus of Stages 6 to 9 exists to manage the position it creates.

Start from what CVA actually is

Credit valuation adjustment (CVA) is the expected loss the bank will suffer if the counterparty defaults, computed today, over the whole life of the portfolio. It is the difference between the value of the derivative assuming the counterparty always pays and its value taking default into account. Equivalently — and this is the more useful framing — it is the price of the credit protection the bank would have to buy to eliminate the risk. If a bank could purchase perfect insurance against Helvetia's default on this swap, the premium would be the CVA.

That framing settles a recurring argument. CVA is not a fee, a buffer, a provision or a cushion. It is a market price for an identified, tradeable risk. It moves when credit spreads move, exactly as an option price moves when volatility moves. Under IFRS 13, fair value must incorporate counterparty credit risk, which makes CVA an accounting number that flows through the profit and loss account, not a prudential add-on that sits quietly on a risk report. Part 1 established that; Step 9 and Step 10 show what it costs.

Three ingredients determine the number, and they multiply.

How much could I be owed when they default? That is the expected exposure profile from the simulation — the mean of the positive part of the future portfolio value, at each future date.

How likely are they to default at each point in the future? That comes from the credit curve, converted into a term structure of default probabilities.

How much do I lose if they do? That is LGD, one minus recovery.

Multiply the three, discount to today, and sum across time.

The formula in words, then in symbols

In words: CVA equals loss given default, multiplied by the sum over all future time buckets of the discounted expected exposure in that bucket, weighted by the probability that default occurs in that bucket.

In symbols, as a discrete sum over time buckets indexed by i, where t is the bucket mid-point:

CVA = LGD × Σ over i of [ EE(t_i) × DF(t_i) × ( S(t_i−1) − S(t_i) ) ]

EE(t) is the expected exposure at time t. DF(t) is the discount factor. S(t) is the survival probability — the probability that the counterparty has not defaulted by time t. The bracketed difference is the marginal probability of default in bucket i: the probability of surviving to the start of the bucket and defaulting within it.

The survival probability comes from the hazard rate, the instantaneous rate of default per unit time for a counterparty that has survived so far. Under the standard credit-triangle approximation, the hazard rate is the credit spread divided by one minus the recovery rate. With a constant hazard rate, S(t) = exp(−hazard × t). Helvetia's ten-year proxy spread is 220bp and its recovery is 40%, so the hazard rate is 0.0220 / 0.60 ≈ 3.67% per annum, and the ten-year survival probability is exp(−0.367) ≈ 69.3%. In other words, the market-implied cumulative probability that Helvetia fails at some point over the swap's life is about 31%. That is the number that surprises people who have anchored on a 0.3% one-year PD.

The useful way to hold CVA in your head is as the premium on an insurance policy the bank is largely forced to self-insure. If a liquid market existed in protection on every client, the bank would simply buy the cover and pass on the premium, and CVA would be a quoted price rather than a calculation. For most corporate counterparties no such market exists, so the bank computes what the premium would be, charges it, and then carries the policy on its own book — which is why the number has to be both a fair price and a defensible reserve at the same time.

Method 1 — the full Monte Carlo calculation

The production method builds CVA from a simulation of the future. It runs in five steps.

Simulate the market. Choose models for the risk factors that drive the portfolio's value — for a cross-currency swap, the EUR and USD interest-rate curves and the EUR/USD spot rate, with their volatilities and correlations. Generate several thousand paths of those factors forward to the trade's maturity, evaluated on a time grid.

Price the portfolio along every path. At each node of every path, revalue every trade in the netting set using the simulated market state at that node. This is the expensive step: it is a full portfolio revaluation performed hundreds of thousands of times.

Net and take the positive part. Sum the trade values within the netting set to get the net portfolio value at that node, then take the greater of that value and zero. The bank loses money on default only where it is owed money; where it owes money it still has to pay.

Average across paths. At each time point, the mean of the exposure across all paths is the expected exposure. The chosen quantile of the same cross-section is the PFE.

Integrate against default probability. Weight the discounted EE profile by the marginal default probabilities from the credit curve, multiply by LGD, and sum.

The result is one number. It is also a number with sensitivities: because it is computed from a model, it can be differentiated with respect to FX spot, interest rates and credit spreads, and those sensitivities are what the XVA desk hedges in Step 9.

Method 2 — the approximations that a salesperson can actually use

A trader quoting a price cannot wait for a Monte Carlo run, so the market uses approximations, and they are better than their reputation.

The swaption-replication view is the more elegant. The exposure of a single interest-rate swap to the bank, at a future date, is the value of the remaining swap when it is positive — which is exactly the payoff of a swaption expiring at that date. A swap's CVA can therefore be approximated as a portfolio of swaptions, one per time bucket, each weighted by the marginal default probability of that bucket and by LGD. This turns a simulation problem into a strip of vanilla option prices that can be read off the volatility surface. For a cross-currency swap the same logic holds with FX options and cross-currency swaptions, with the extra complication that the terminal notional exchange contributes a large, long-dated FX option component.

The rule of thumb is cruder and far more widely used: CVA is approximately EPE multiplied by the credit spread multiplied by the risky duration, where EPE is the expected positive exposure and risky duration is the discounted, survival-weighted duration of the trade. For Helvetia: EPE is EUR 22.4m, the ten-year spread is 220bp, and the risky duration with a 3.67% hazard rate and a EUR discount curve around 1.5% is about 7.8 years. That gives 22.4 × 0.0220 × 7.8 ≈ EUR 3.85m, against a full-calculation answer of EUR 4.2m. The approximation is roughly 8% low.

An approximation that is eight per cent wrong is extremely useful, and understanding why is a lesson in what pricing infrastructure is for. The salesperson does not need the exact number; they need to know whether the charge is four million or forty million, because that determines whether the conversation with the client is about the last basis point or about whether the trade happens at all. The approximation is also stable and explicable: a trader can see immediately that halving the tenor roughly halves the charge, and can defend the logic to a client. The full calculation then runs before the trade is confirmed, and the difference between the two is a P&L item the desk owns rather than a surprise. What would be unacceptable is an approximation that is wrong in an unknown direction by an unknown amount, which is what happens when banks approximate without ever reconciling to the full run.

The worked calculation for Helvetia

The table below reconstructs the CVA from the ingredients, in five time buckets. Hazard rates are bootstrapped from the two points on the proxy curve: the five-year spread of 180bp implies a hazard of 0.0180 / 0.60 = 3.00% for years zero to five, and the ten-year spread of 220bp implies an average hazard of 3.67% over ten years, so the forward hazard for years five to ten is about 4.33%. Survival probabilities follow from S(t) = exp(−hazard × t) applied piecewise. Discount factors use a flat 1.5% EUR continuous rate. LGD is 60%.

Bucket EE at mid-point (EUR m) Discount factor Marginal PD CVA contribution (EUR m)
Years 0–2 15.0 0.985 5.82% 0.52
Years 2–4 24.0 0.956 5.49% 0.76
Years 4–6 30.0 0.928 6.27% 1.05
Years 6–8 28.0 0.900 6.83% 1.03
Years 8–10 25.5 0.874 6.28% 0.84
Total EPE 22.4 on the monthly grid — 30.7% cumulative 4.20

Each contribution is LGD multiplied by EE, by the discount factor, and by the marginal PD. The first row: 0.60 × 15.0 × 0.985 × 0.0582 = EUR 0.52m. The total is EUR 4.2m, which is 91bp of the EUR 460m notional, or roughly 11bp per annum running when spread over the risky duration.

Two features of the table are worth reading carefully. The largest contributions come from the middle and late buckets, not the early ones, because for a cross-currency swap the exposure keeps growing as FX drifts away from the strike and the terminal notional exchange stays in front of you. Compare an amortising loan, where the biggest loss exposure is at the start. And the marginal default probabilities rise in the later buckets despite survival falling, because the bootstrapped forward hazard rate is higher beyond year five. Those two effects compound, which is why long-dated uncollateralised trades with weaker credits are so much more expensive than a linear intuition suggests.

Note that the arithmetic mean of the five bucket EEs, about EUR 24.6m, sits above the reported EPE of EUR 22.4m. That is a bucketing artefact: five two-year buckets sampled at their mid-points over-weight the fat centre of a concave profile, where the production run uses a monthly grid. Reconciling a coarse desk calculation to the production number is a routine and instructive model-validation exercise, and the gap being explicable is the point. All figures illustrative only.

Figure 4 CVA triangulation: four inputs, one price. Illustrative figures.

The rest of the XVA family

CVA was the first adjustment and remains the largest, but it is not the only one. The valuation of an uncollateralised derivative differs from its textbook value for several distinct reasons, and the market has given each its own name.

DVA — debit valuation adjustment is CVA seen from the other side. Just as the bank might lose money if the counterparty defaults, the counterparty might lose money if the bank defaults — and from the bank's own accounting perspective, that potential loss to the counterparty is a gain to the bank. DVA is therefore a positive adjustment to the bank's own income that increases as the bank's own credit deteriorates. The discomfort is real and the arithmetic is correct. A bank whose spread widens by 100bp books a DVA gain, which is to say it reports profit because the market has decided it is more likely to fail. Regulators addressed this by requiring DVA gains and losses on a bank's own credit to be filtered out of regulatory capital, so the accounting number survives while its capital effect does not. Whether DVA should be hedged is a longer argument: hedging it means selling protection on yourself, which is not straightforwardly available.

FVA — funding valuation adjustment prices the cost of funding an uncollateralised derivative. If the bank hedges the Helvetia swap in the interbank market under a CSA, it will have to post variation margin on the hedge while receiving nothing on the client trade. When the hedge moves against the bank, it must fund cash margin at its own borrowing spread. FVA is the present value of that expected funding cost, and it is a cost to the bank on the asset side of the profile.

MVA — margin valuation adjustment prices the cost of posting initial margin over the life of a trade. Initial margin is segregated and cannot be rehypothecated, so the poster funds it and earns nothing useful on it. MVA is the present value of that funding drag. It became material only after the Uncleared Margin Rules made initial margin near-universal for large market participants, and it is the reason a cleared trade and an uncleared trade with identical economics have different prices.

KVA — capital valuation adjustment prices the cost of the regulatory capital the trade consumes over its life. Capital is not free: shareholders require a return on it, and a trade that locks up capital for ten years must earn that return or it destroys value. KVA is the present value of the cost of holding the required capital over the trade's life, at the bank's hurdle rate. It is the least standardised of the family, because the capital rules themselves change.

ColVA — collateral valuation adjustment prices the optionality and the economics embedded in the collateral agreement itself. A CSA that permits several currencies of cash collateral gives the poster a cheapest-to-deliver option, because the interest rate paid on each currency differs and the poster will choose the cheapest. A CSA that permits securities gives a wider option still. ColVA values that optionality and any mismatch between the CSA's remuneration terms and the bank's actual funding.

Adjustment What it prices Who owns it Sign for the bank
CVA Expected loss from counterparty default XVA desk Cost
DVA Expected gain from the bank's own default Accounting; filtered from capital Gain
FVA Cost of funding uncollateralised positions XVA or treasury Cost
MVA Cost of funding initial margin XVA desk Cost
KVA Cost of regulatory capital consumed Capital management Cost
ColVA Value of collateral optionality and CSA terms XVA or collateral desk Either

The DVA-FVA overlap is the sharpest unresolved argument in derivative valuation, and it is worth understanding rather than memorising.

DVA says: because the bank may default, it will not have to pay some of what it owes, and that saving is a gain today. FVA says: because the bank funds unsecured at a spread over the risk-free rate, financing an uncollateralised derivative costs money, and that cost is a charge today. The problem is that the bank's unsecured funding spread is, to a first approximation, compensation to its lenders for exactly the default risk that DVA monetises. Charging FVA in full and booking DVA in full therefore counts the same credit spread twice, once as a cost and once as a benefit.

The theoretical resolution, associated with the Hull-White critique, is that the funding benefit of an uncollateralised liability is DVA, so a symmetric framework should not add a separate FVA on top. The practical resolution, which the industry adopted after several banks took large one-off FVA charges in 2013 and 2014, is that FVA reflects an observable cost the treasury actually incurs and that traders are actually charged, whereas DVA reflects a benefit that only accrues in default and cannot be monetised beforehand. Banks therefore run both, with the overlap acknowledged and managed rather than eliminated.

The practical takeaway for anyone building or reviewing an XVA framework: check what discount curve and what funding spread each adjustment uses, and check whether the same spread appears in two places with opposite signs. It usually does, and whether that is right depends on a convention the bank should be able to state explicitly.

How the charge actually reaches the client

The mechanism matters as much as the number, because it determines who bears the risk of getting it wrong.

The XVA desk is a central unit that owns the counterparty-risk position of the whole bank. When a sales team wants to do the Helvetia trade, it requests an XVA quote. The XVA desk prices the incremental CVA, FVA and KVA of the trade against the existing netting set and returns a charge — EUR 6.2m in total for Helvetia. The rates desk then prices the swap itself at market, adds the XVA charge, and the salesperson quotes the client the all-in price.

When the trade books, the XVA charge is paid internally: the rates desk transfers EUR 6.2m of the day-one revenue to the XVA desk. The rates desk keeps a clean market-risk position with no credit component. The XVA desk holds the counterparty-risk position, the reserve, and the obligation to hedge it. Every subsequent move in Helvetia's credit spread hits the XVA desk's P&L, not the rates desk's.

This internal transfer is what stops the sales desk from writing free options on credit. Without it, a salesperson could win a mandate by quoting an aggressive price, book the revenue, and leave the credit risk sitting somewhere unowned until it turned into a loss years later, by which time the bonus had been paid and the salesperson had moved on. With it, the price of the credit risk is extracted from the deal revenue on day one, and any decision to under-charge shows up immediately as a smaller day-one P&L for the desk that made the decision. Aligning the incentive to the moment of the decision is the whole design.

The XVA desk, in turn, is not a profit centre in the usual sense. Its mandate is to be flat: to charge the right price, hedge what can be hedged, and report the residual. A large XVA desk profit generally means the bank has been over-charging clients or has taken an unhedged directional credit view, and both should be visible.

  1. In one sentence, what is CVA?

  2. Why does the calculation need a credit spread rather than a rating?

  3. Helvetia has no traded CDS. What does the bank do instead?

Step 6 — The dullest stage in the lifecycle is where most counterparty-risk failures begin

Booking a trade is data entry, and data entry is where the risk model is silently corrupted. Nobody writes a case study about trade capture. Every serious counterparty-risk incident review, however, contains at least one finding that begins with the words "the trade was booked against the wrong".

Five attributes are set at booking, and every downstream number inherits all five without question.

The netting set identifier determines which trades are netted against which. Book the Helvetia swap into the wrong netting set and the exposure engine will offset it against positions belonging to a different counterparty, producing an exposure number that is not merely inaccurate but structurally meaningless. Book it into a new netting set when it should have joined an existing one and the bank forgoes real netting benefit, over-states exposure, over-charges the client and over-consumes capital.

The CSA link determines whether the exposure engine applies collateral. A trade booked as collateralised when there is no CSA produces an exposure profile collapsed to a few days of market move rather than the true ten-year profile. The reported exposure will be a small fraction of the real one, the CVA will be a small fraction of the real one, and nothing will look wrong on any report.

The legal entity determines which netting opinion, which jurisdiction and which group the trade rolls up to. This is the Stage 1 problem arriving in operational form.

The maturity drives the tenor bucket, the limit consumed and the capital add-on. A date keyed as 2026 rather than 2036 removes the trade from every long-tenor control.

The product classification determines which model prices it, which supervisory factor applies under SA-CCR, and which regulatory report it appears in. Classify a cross-currency swap as an interest-rate swap and the FX component vanishes from the risk view.

None of these five errors produces an error message. Each produces a plausible number that is wrong, and a plausible number that is wrong is worse than a missing one, because it will be used.

Confirmation and affirmation are the counterparty's check on your data

A confirmation is the legal document recording the economic terms of a trade, exchanged and agreed between the two parties. Affirmation is the faster electronic step of agreeing the terms before the formal confirmation is issued. Both are, in effect, an external reconciliation: the counterparty independently records the same trade and the two records are compared. Unconfirmed trades are the population where the bank's version and the client's version may differ, and where a dispute at close-out becomes likely rather than possible. Regulators require the volume and ageing of unconfirmed trades to be reported precisely because it is a good leading indicator of operational quality.

T+0 discipline — capturing, validating and confirming on trade date — matters for a reason specific to counterparty risk. The overnight exposure batch runs on the trade population as at the close. A trade not captured by the cut-off is invisible to the entire risk, limit, collateral and capital chain until the next night, which means the bank spends a day without knowing what it owns. In a fast market, that day is the one that matters.

Reconciliation is the control that catches what validation misses

Front office and risk systems hold separate representations of the same portfolio. A daily reconciliation compares them on trade count, notional, maturity, product and value, and investigates the breaks. Separately, portfolio reconciliation with the counterparty — mandated under EMIR for non-cleared derivatives, at a frequency scaled to portfolio size — compares the bank's population and valuations against the client's. Trades in one book and not the other, and valuations that differ beyond tolerance, are the raw material of margin disputes in Step 8 and close-out litigation in Step 10.

Trade capture is a data-quality problem with a legal deadline. Three controls carry most of the weight.

Validation at the point of entry, not downstream: reject a booking where the counterparty entity is not fully onboarded, where the netting-set identifier does not exist, or where the CSA flag contradicts the documentation system's record. The documentation system, not the trader's screen, is the authority on whether a CSA exists.

Golden-source discipline for the netting set and CSA link: one system owns the mapping, everything else reads it. Where the trading system, risk system and collateral system each derive the netting set from their own logic, they will diverge, and the divergence will be discovered during a default.

Full lineage from the booked trade to every downstream number, so that when an exposure figure is challenged the bank can trace it back through aggregation, simulation and capture to the executed confirmation. Part 4 develops this.

Step 7 — Exposure is a distribution, and the nightly batch is the machine that produces it

This is the analytical heart of counterparty risk and the largest piece of infrastructure in the chain. Everything upstream prepares for it and everything downstream consumes it. It runs every night, on every netting set, and its output is not a number but a family of profiles from which every control in the business takes a different reading.

The vocabulary, defined precisely, because these terms are used loosely and should not be

Mark-to-market (MtM) is the current value of the netting set: the sum of the values of all trades in it, positive and negative. It can be either sign. It is what the portfolio is worth today.

Exposure is the greater of the mark-to-market and zero. This asymmetry is the definition of credit risk and it is worth pausing on. If the counterparty defaults when the netting set is worth EUR 40m to the bank, the bank has lost EUR 40m of value and must file a claim. If the counterparty defaults when the netting set is worth negative EUR 40m to the bank, the bank still owes EUR 40m and must pay it — close-out does not release it from its own obligations. The bank has upside risk on the counterparty's default and no downside benefit. That is the option-like payoff that makes CVA a derivative.

Expected exposure (EE) at a future date is the average of the exposure across all simulated states of the world at that date. It is a function of time, not a single number: EE at one year, EE at two years, and so on, producing the profile.

Expected positive exposure (EPE) is the time-average of the EE profile over a stated horizon. It collapses the profile to a single number and is the natural input to a CVA rule of thumb and to capital formulas.

Effective EPE (EEPE) is a regulatory refinement. Effective EE at each date is defined as the running maximum of EE up to that date, so it never decreases. EEPE is the average of effective EE over the first year. The non-decreasing construction exists to stop short-dated portfolios that roll over from showing a declining profile and thereby understating capital, since in practice the trades are replaced. It is a capital measure, and it is used nowhere else.

Potential future exposure (PFE) at a future date is a high quantile of the exposure distribution at that date — the 95th, 97.5th or 99th percentile depending on policy. Peak PFE is the maximum of the PFE profile across all future dates. It is the limit measure.

Measure What question it answers Where it is used
MtM What is the portfolio worth today? Collateral calls, accounting, daily P&L
Exposure What would I lose if they defaulted right now? Current exposure reporting, SA-CCR replacement cost
EE at a future date What do I expect to be owed on that date? The CVA integral
EPE What do I expect to be owed on average over the life? CVA rule of thumb, pricing conversations
EEPE What is the regulatory average over the first year? IMM capital calculation
PFE and peak PFE What is the worst I could plausibly be owed? Limit setting and monitoring
Figure 5 The exposure profile of the Helvetia swap: expected exposure, EPE and the PFE limit. Illustrative figures.

A loan exposure is a level; a derivative exposure is a forecast. A lender knows the balance outstanding today and knows the schedule that reduces it, so the only real uncertainty is whether it gets paid. A swap desk knows neither, because what it will be owed in four years depends on where EUR/USD and two yield curves have travelled — and the honest answer is a distribution with a shape rather than a number with a value. Everything technical in this stage follows from that one difference.

Why the profile has the shape it has

Two forces act on the exposure profile, in opposite directions, and the shape is the outcome of the race between them.

The first is diffusion. As you look further into the future, the range of possible market states widens. FX rates, interest rates and volatilities wander further from today's values as time passes, roughly with the square root of time. A wider distribution of market states means a wider distribution of portfolio values, and since exposure is the positive part, a wider distribution means a higher mean and a much higher upper quantile. Diffusion pushes exposure up over time.

The second is amortisation and the passage of cashflows. As the trade ages, payments are made and the remaining contract shrinks. A five-year swap that has run for four years has one year of cashflows left and can be worth much less than it could have been at inception. Roll-down pushes exposure down over time.

For a vanilla interest-rate swap the two forces produce the characteristic hump. Early on, diffusion dominates because the remaining life is long, so exposure rises. Later, roll-down dominates because there is little left to be worth anything, so exposure falls, reaching zero at maturity. The peak sits somewhere between a third and a half of the way through the life.

For a cross-currency swap the shape is different and the difference is the whole point of choosing this trade as the running example. A fixed-for-fixed cross-currency swap exchanges principal at maturity: the bank delivers EUR 460m and receives USD 500m on 15 March 2036. That final exchange is a ten-year FX forward, and its value does not roll down. Right up to the maturity date, the full EUR 460m against USD 500m remains at risk to the EUR/USD rate. Diffusion continues to widen the distribution of that final exchange while there is nothing to amortise, so the tail stays fat until the last day.

The practical consequences are large. The peak PFE arrives later and is higher than for an equivalent-notional interest-rate swap. Roughly two-thirds of the CVA sits beyond year four, as the worked table in Step 5 shows. And the trade is at its most dangerous when it is nearly over — the opposite of the intuition most people bring from lending, where the last instalment on a loan is the smallest risk.

The Monte Carlo machinery, and why it is expensive

Simulating the future is conceptually simple and computationally brutal. The engine has four components.

Risk-factor evolution models. Each risk factor gets a stochastic process. Interest rates commonly use a one- or two-factor short-rate model, or a market model where the portfolio demands it. FX spot commonly uses geometric Brownian motion with a stochastic or term-structure volatility. Equity and commodity factors get their own. The factors are correlated, and the correlation matrix is calibrated from history. Two calibration choices matter enormously and are frequently under-examined: whether the model is calibrated to historical dynamics or to market-implied volatilities, and how correlations are estimated in a way that keeps the matrix positive-definite when it has thousands of entries.

The number of paths. Typically two thousand to ten thousand. This is far fewer than a derivative-pricing Monte Carlo would use, and the reason is the tail. Two thousand paths give a stable mean and therefore a decent EE, but a 99th-percentile PFE estimated from two thousand paths rests on the twenty worst of them, and its sampling error is large. Banks that set limits at 99% on two thousand paths are measuring noise at the very point where they are supposed to be most careful.

The time grid. Dense early, sparse late. Daily or weekly for the first month, monthly for the first year or two, quarterly then annually thereafter. The reason is that the early grid must resolve collateral dynamics and near-term margin calls, while the far end only needs to resolve the shape of the profile. A typical grid for a ten-year trade has sixty to a hundred and twenty points.

Pricing along the paths. At every node the portfolio must be revalued. Analytic formulas are used where available; where a trade needs its own nested simulation to price, that trade will dominate the runtime, and banks use regression-based approximations or pre-computed pricing grids to avoid the nested Monte Carlo.

The cost scales as paths multiplied by time steps multiplied by trades. Five thousand paths, one hundred time steps and one thousand trades in a netting set is five hundred million trade valuations for one counterparty. Multiply across several thousand counterparties and the batch is in the hundreds of billions of valuations. This is why exposure computation was the last major bank calculation to move off overnight batch, why adjoint algorithmic differentiation matters so much for producing sensitivities cheaply, and why GPU acceleration has changed the economics of this stage more than any regulatory change. Part 5 develops the transformation.

The same word, two measures — and most reconciliation arguments start here

The bank computes exposure twice, under two different probability measures, for two different purposes, and calls both of them exposure. This is the single most common source of confusion in the discipline, and it is not a subtlety. Potential future exposure feeds limits, and a limit is a forecast: the question is what this portfolio will plausibly be worth, so the simulation is calibrated to how markets have actually behaved, using historical volatilities and historical drift. That is the real-world measure. CVA feeds a price, and a price is the cost of a hedge: the question is what the market would charge today to take this risk, so the simulation must be calibrated to what the market is quoting, using volatilities implied by traded options and drift set by forward rates. That is the risk-neutral measure.

The two calibrations give different answers for the same portfolio, and both are correct. Implied volatility normally sits above realised volatility, because it contains a risk premium that buyers of options pay. Feed implied volatility into a PFE calculation and the tail widens for a reason that has nothing to do with what is likely to happen; feed historical volatility into a CVA calculation and the charge comes out below what the hedge would actually cost. Many banks therefore run two engines, or one engine with two calibration sets, and produce two exposure profiles that disagree by design.

Knowing this saves a great deal of pointless argument. When the risk function and the XVA desk produce different exposure numbers for the same trade, the first question is not whose model is wrong. It is which measure each number was computed under, and a large share of these disputes end there.

The netting-set aggregation step is where the risk mitigation actually happens

The order of operations in the simulation is not a detail; it is the difference between a right answer and a wrong one. At each node of each path, the engine must first sum the values of all trades in the netting set, and only then take the positive part. Applying the positive-part operation trade by trade and then summing gives the gross figure — mathematically the sum of maxima, which is always at least as large as the maximum of the sum.

For an active netting set the difference is enormous. The canonical illustration in this module uses a whole-portfolio view of one counterparty: gross positive mark-to-market of EUR 142m collapses to EUR 38m after close-out netting. That EUR 104m of reduction is created entirely by one line of code sitting on one clean legal opinion. Get the netting flag wrong in Stage 2 or the netting set identifier wrong in Stage 6, and the bank either reports EUR 142m of exposure it does not have or EUR 38m of exposure it cannot legally claim.

A netting set is a bucket, and the contract defines the bucket, not the trading desk. Trades inside the same bucket offset each other on default; trades in different buckets do not, however similar they look economically. Two swaps that hedge one another perfectly but sit under different Master Agreements are, for every purpose in this module, two separate risks that cancel on a spreadsheet and not in a bankruptcy court.

Collateral inside the simulation, and why a CSA never gets you to zero

Where a CSA exists, the simulation must model it, because collateral changes the exposure at every node. At each node the engine computes the netting set's value, works out what collateral would have been called and delivered given the threshold, the MTA and the call frequency, and reduces the exposure by the collateral held. Done properly, this collapses the exposure profile dramatically — but not to zero, and the residual has a name.

The margin period of risk (MPoR) is the assumed time between the last effective exchange of collateral and the completion of close-out. It covers the sequence that actually happens: the counterparty misses a margin call, the bank chases it, a grace period runs, the bank declares an event of default, designates an Early Termination Date, values the portfolio and replaces or unwinds the hedges. During that whole window the market keeps moving while the collateral is frozen at its last level.

The supervisory assumptions are ten business days for uncleared derivatives and five business days for cleared — the shorter cleared figure reflecting a CCP's ability to act faster, with pre-agreed default-management procedures and standing auction arrangements. Both are extended in defined circumstances: for netting sets with a large number of trades, for illiquid collateral, and for netting sets that have experienced repeated margin disputes.

The MPoR is why a fully collateralised netting set still has exposure, and it is the single most important parameter in collateralised exposure modelling. Because exposure over the MPoR grows roughly with the square root of the window, moving from five to ten days increases the residual by about forty per cent. In the illustrative whole-portfolio cascade, EUR 38m of netted exposure falls to EUR 12m after variation margin, not to zero. That EUR 12m is essentially ten days of market move on a portfolio whose collateral is stale, plus the threshold and MTA that the CSA permits to go uncollateralised, plus disputed amounts.

The MPoR is where collateral modelling gets genuinely difficult, and where most banks make a simplification worth knowing about.

The rigorous treatment simulates the collateral balance path by path with the correct lag: at any point in time, the collateral held reflects the portfolio value as it stood one margin period of risk earlier, not as it stands now. That requires the engine to carry a lagged state variable through the simulation and to model the cashflows that fall inside the MPoR window, which are paid by the bank but may not be received from the defaulting counterparty. The common simplification is to compute uncollateralised exposure and apply a closed-form or grid-based collateral adjustment afterwards. It is faster and it systematically understates the tail, because it misses the correlation between a large market move, a large margin call and a failure to deliver.

The second-order effects that the simplification also misses are the ones that showed up in 2008: the counterparty stops delivering collateral before it formally defaults, disputes the call to buy time, and delivers the cheapest and least liquid eligible asset in the days beforehand. None of these appear in a model that assumes collateral is exchanged mechanically until the moment of default.

The nightly cycle has six steps and a twelve-hour window

The daily exposure cycle is a defined operational sequence, running roughly from 18:00 to 06:00, and each step has a downstream consumer that fails if it fails.

Step Window What is produced What breaks if it fails
1. Market data close and curve build 18:00–20:00 Validated closing rates, bootstrapped curves, vol surfaces, credit curves Everything; the batch runs on wrong prices or not at all
2. Trade population snapshot 19:00–21:00 Frozen, reconciled list of live trades by netting set with CSA links Missing or duplicated trades; exposure silently wrong
3. Monte Carlo path generation 21:00–23:00 Simulated risk-factor paths on the time grid No simulation; fallback to the previous day's profiles
4. Pricing along paths and aggregation 23:00–04:00 Netted exposure at every node; EE, EPE and PFE profiles The core output is missing; limits run on stale numbers
5. Collateral and limit application 04:00–05:00 Collateral-adjusted profiles, limit utilisation, breach list Breaches undetected; traders start the day unconstrained
6. Distribution to downstream systems 05:00–06:00 Feeds to limit, capital, collateral, XVA and reporting systems Reports and regulatory submissions late or inconsistent

The batch has to finish before the first trading desk opens, and that constraint has shaped bank architecture for twenty years. When it overruns, the bank starts the day on yesterday's exposure. When it fails, there is rarely time to rerun it. The standard mitigations are a documented fallback to the prior day's profiles with an explicit flag on every downstream report, a restartable design so a failure at step four does not force a rerun from step one, and a critical-path model that identifies which netting sets must complete before the others.

The most common failure is not a crash; it is a silent completion on bad data. Step one produces a curve from a stale quote, the batch runs perfectly, and every number downstream is wrong with no alarm. This is why market-data validation — tolerance checks against the prior close, cross-checks between related instruments, stale-quote detection — is the most valuable control in the entire cycle and the one most often treated as a technical detail.

Figure 6 The nightly cycle: twelve hours, one deadline, no second chance. Illustrative timings.

The shape of the night matters as much as its length. Nearly half the window goes to one step — pricing along paths and aggregating into netting sets — and that step is the critical path in the strict project-management sense: nothing downstream can start until it finishes, and it cannot be parallelised beyond the grid it already runs on. The steps around it are cheap by comparison. When a bank says it needs more compute for counterparty risk, this is almost always the step it means.

The deadline is not a service-level target, it is a fact about the trading day. At six in the morning the desks open, the limit system has to know what capacity exists, the collateral team has to know what to call for, and the capital number has to be defensible. A cycle that overruns does not degrade gracefully. The bank simply opens with yesterday's exposure, which means limits are enforced against stale utilisation, margin calls are missed or wrong, and nobody can say with confidence what the institution is owed. That is why the contingency buffer at the end of the window is not slack to be optimised away — it is the difference between a late batch and a day without controls.

Wrong-way risk is the failure of the independence assumption

The CVA formula multiplies exposure by default probability as though the two were independent. Often they are not, and where they are not the formula understates the loss, sometimes by a lot.

General wrong-way risk arises where the counterparty's probability of default is correlated with the market factors that drive the bank's exposure, through a broad economic relationship rather than a direct contractual one. Three examples make the pattern clear.

An oil producer hedging its oil output sells forward. When oil prices fall, the hedge becomes valuable to the producer and the bank is out of the money, so there is no exposure. But when oil prices rise sharply, the hedge is deeply out of the money for the producer and deeply in the money for the bank — and the producer, having locked in low prices while its costs and its competitors' revenues rise, may be under severe strain. The bank's exposure is largest precisely when the counterparty is weakest.

A bank buying CDS protection on a sovereign from a bank domiciled in that sovereign is the classic case, and the one worth remembering because it is so stark. Protection on Italy bought from an Italian bank pays out only when Italy defaults, which is the state of the world in which an Italian bank is least likely to be able to pay. The hedge is worth most exactly when it is worth nothing. This construction was widespread before 2008 and is now the standard illustration of the concept in supervisory guidance.

An emerging-market corporate borrowing in dollars and swapping into local currency is exposed to a devaluation, and a devaluation is also what makes the bank's swap position valuable. The exposure and the default probability are driven by the same variable in the same direction.

Specific wrong-way risk is narrower, more severe and separately capitalised. It arises where the exposure is directly linked to the counterparty's own credit quality by the structure of the transaction. Taking a client's own shares as collateral is the textbook case: in a default the collateral is worthless at exactly the moment it is needed. A repo in which the collateral is a bond issued by the counterparty or an affiliate is the same construction. A total return swap referencing the counterparty's own equity is another. Supervisors require specific wrong-way risk to be identified transaction by transaction and treated with a punitive exposure calculation, because no correlation parameter can capture a relationship this direct.

The modelling responses vary in ambition. The simplest is a multiplier on the exposure for portfolios flagged as wrong-way. More sophisticated is to model the counterparty's default intensity as a stochastic process correlated with the market factors in the same simulation, which produces a correct answer at the cost of calibrating a correlation that is almost never observable. Most banks land on a mix: identify wrong-way exposures qualitatively, apply a conservative adjustment, and prohibit the specific cases outright. The identification step is worth more than the modelling step, and it is a credit-officer judgement, not a quant output.

Figure 7 Wrong-way risk: exposure and default probability rising together. Illustrative figures.

The arithmetic of wrong-way risk is worse than the intuition suggests, because two numbers are multiplied rather than added. CVA is exposure multiplied by default probability. If the two are independent, the average of the product is the product of the averages and the standard calculation holds. If they move together, the average of the product exceeds it, and the shortfall grows with the correlation. On the illustrative Helvetia figures, a CVA of EUR 4.2m computed under independence becomes EUR 6.8m at a 30% correlation — a sixty per cent increase produced entirely by the joint behaviour, with no change to either the exposure profile or the credit curve taken on its own.

Right-way risk is the mirror image, and the framework treats it asymmetrically

Exposure and credit quality can also move in opposite directions, and when they do the risk is genuinely smaller. A bank that buys commodity price protection from a producer is in this position: if the commodity price falls the bank is owed money, but the producer that owes it is a buyer's beneficiary rather than a victim, and its credit is not deteriorating in the state of the world where the exposure exists. Airlines hedging fuel, importers hedging a currency they are short of, and exporters hedging a currency they are long of all generate right-way structures when the hedge is set up in the natural direction.

The regulatory framework penalises wrong-way risk far more readily than it credits right-way risk, and that asymmetry is deliberate. Specific wrong-way risk attracts an explicit and punitive treatment. Right-way risk earns, at best, the absence of a penalty. The supervisory logic is that a correlation assumption which reduces capital is a correlation assumption that will be tested in exactly the conditions where it is least likely to hold, so the burden of proof runs one way only. For a bank building the model this matters commercially: modelling wrong-way risk properly is a cost that must be borne, and modelling right-way risk properly buys very little relief.

Backtesting is how you find out whether the model was ever right

A simulation of the future can only be validated against the future. Backtesting compares the model's predicted distribution against what actually happened: for each historical date, the model predicted a distribution of portfolio values one month, three months and one year ahead; the realised value can be located in that predicted distribution; and across many dates and many netting sets the realisations should be uniformly spread across the quantiles. If realised values land above the 99th percentile far more often than one per cent of the time, the model understates the tail.

Backtesting the exposure model is a formal requirement for any bank using the Internal Model Method, and supervisors specify the counterparty portfolios, horizons and statistical tests. Even where a bank uses only the standardised approach for capital, it should backtest, because the same model sets its limits and its CVA. A model that has never been confronted with an outcome is an assumption with a user interface.

Model validation extends further: independent review of the model's theory and implementation, benchmarking against alternative models, sensitivity testing of key assumptions such as correlations and mean-reversion parameters, and stress testing against historical episodes. The validation function must be independent of the model developers and must have the authority to restrict the model's use. That authority is the part that is most often nominal.

The exposure engine is the largest compute consumer in a CIB risk stack and the piece most transformed by the last five years of infrastructure change.

The traditional design is a scheduled overnight batch on an on-premise grid, writing profiles to a warehouse that limits, capital and reporting read the next morning. The emerging design decouples the pieces: risk-factor simulation as a service, pricing distributed across GPU nodes, results streamed into a store that supports both the overnight snapshot and intraday incremental updates. The business driver is not elegance. It is that a bank able to recompute exposure intraday can manage a stressed counterparty on the day rather than the morning after.

Two techniques do most of the work. Adjoint algorithmic differentiation (AAD) computes all the sensitivities of a valuation in roughly the cost of a few valuations, rather than one re-run per risk factor, which turns hedging an XVA book from computationally impossible into routine. And American Monte Carlo regression techniques let the engine value path-dependent and early-exercise trades along simulated paths without nesting a second simulation inside the first.

The data requirement is unglamorous and decisive: a validated, complete, point-in-time market-data set every night, with full lineage. Most exposure-engine incidents are market-data incidents wearing a modelling costume.

Step 8 — Collateral converts credit risk into liquidity risk; it does not delete risk

Collateral is the most effective counterparty-risk mitigant ever devised and it is the reason the derivative market survived 2008. It is also the mechanism by which two of the three worst market episodes since then were transmitted. Both statements are true, and holding them together is the whole of Stage 8.

Variation margin settles the past

Variation margin (VM) is the daily transfer of cash equal to the change in the mark-to-market of the netting set. If the portfolio moves EUR 5m in the bank's favour today, the counterparty delivers EUR 5m tomorrow. If it moves back, the bank returns it. The mechanics are the CSA's: value the portfolio, compare to collateral already held, apply the threshold and the MTA, issue a call for the difference.

The economic effect is to reset the exposure to approximately zero every day. What remains is not zero, and it consists of exactly three things. The threshold, the unsecured amount the CSA permits before any call is due. The MTA, the amount below which no transfer is made, which floats as a small permanent uncollateralised balance. And the market move over the margin period of risk, the days between the last effective collateral exchange and the completion of close-out, during which the portfolio moves and the collateral does not.

With a zero threshold, a small MTA and daily calls, ten years of credit exposure becomes about ten business days of market exposure. For most portfolios that is a reduction of ninety per cent or more, and it is the reason a collateralised interbank swap carries almost no CVA while the identical trade with Helvetia carries EUR 4.2m.

The margin period of risk is the length of the fuse, not the size of the charge. It measures how long the bank is flying blind between the last collateral it actually received and the moment it has replaced the trade in the market. Every parameter in an initial margin calculation is ultimately an answer to one question: how far can this portfolio move while nobody is topping up the collateral.

Initial margin covers the gap that variation margin leaves

Initial margin (IM) is collateral posted at the outset and held throughout the life of the relationship, sized to cover the potential change in the portfolio's value during the margin period of risk. It is precisely the answer to the residual that VM cannot address: the loss that accrues between the last VM payment received and the completion of close-out.

Two properties distinguish it from variation margin, and both are deliberate. IM is posted by both parties, gross, each to the other, so it is not netted and both sides tie up collateral simultaneously. And IM must be segregated with a third-party custodian and cannot be rehypothecated. The receiver holds a security interest but cannot use the asset. That prohibition is what makes IM genuinely protective: in a default, the collateral is there rather than lent onward through three counterparties.

The industry standard model for calculating IM on uncleared derivatives is ISDA SIMM, the Standard Initial Margin Model. It is a sensitivity-based framework: each party computes the portfolio's sensitivities to a defined set of risk factors, applies prescribed risk weights and correlations, and aggregates to a margin figure. The reason a standard model exists at all is disputes. If each party used its own model, every IM call would be a negotiation. SIMM lets both parties compute the same number from the same sensitivities, which makes the calls mechanical and confines disputes to the sensitivity inputs.

IM is expensive, and understanding why explains MVA. The poster funds the collateral, cannot use it, and receives at best a modest return on it. That funding cost, over the life of the trade, is the margin valuation adjustment. It is a real, ongoing, unavoidable cost of doing business under margin rules, and it is why the total price of a trade depends on the collateral regime as much as on the market.

UMR phased in over six years and finished in September 2022

The Uncleared Margin Rules implement the BCBS-IOSCO framework for margin on non-centrally-cleared derivatives. Variation margin requirements applied broadly and early. Initial margin was phased in from 2016, in annual waves, each catching a smaller cohort defined by AANA, the aggregate average notional amount, being the group-level average notional of non-cleared derivatives measured over a defined observation period.

The thresholds descended wave by wave, and the final phase in September 2022 caught counterparties with AANA above EUR 8bn. That wave was by far the largest by count, bringing several thousand smaller asset managers, funds and corporates into scope for the first time, and it drove a substantial build-out of custody, SIMM calculation and collateral-operations capability across the market.

Helvetia sits below every one of these thresholds and outside the scope entirely, and this is the design of the teaching case. As an NFC−, it is exempt from the clearing obligation and from the margin requirements. It posts nothing. The full ten-year exposure profile sits on the bank's balance sheet, unmitigated, which is why this trade generates EUR 4.2m of CVA where a cleared or margined equivalent would generate almost none.

The CSA parameters in operation

The parameters negotiated in Stage 2 now have daily operational consequences.

The threshold is the standing unsecured line. A EUR 10m threshold means the bank absorbs the first EUR 10m of adverse move permanently, and it appears directly in the CVA.

The MTA avoids trivial transfers. A EUR 500,000 MTA means a call is only made when the uncollateralised amount exceeds the threshold by that much, leaving a small floating unsecured balance at all times.

The eligible collateral schedule lists what may be delivered — typically cash in specified currencies, government bonds of specified issuers and maturities, sometimes supranational or high-grade corporate paper. Haircuts discount non-cash collateral to allow for its price falling during the close-out period. A ten-year government bond might carry a 4% haircut; an equity, if eligible at all, 15% or more.

The currency of collateral creates a valuable option that is easy to give away by accident. A CSA permitting cash in EUR, USD, GBP and JPY lets the poster deliver whichever currency is cheapest to fund on any given day, because the CSA pays the relevant overnight rate on each currency and the poster's own funding cost in each differs. That cheapest-to-deliver option is genuinely worth money, it is exercised continuously, and its value moves with cross-currency basis. Pricing it is ColVA. It also means the discount curve used to value a collateralised derivative depends on the CSA's collateral terms, because the discounting must reflect the actual collateral that will be posted, which is why the same trade under two different CSAs has two different fair values.

Figure 8 The cascade: gross exposure, close-out netting, variation margin, initial margin. Illustrative figures.

The cascade in Figure 8 uses the module's illustrative whole-portfolio view of one counterparty. Gross positive mark-to-market across all trades is EUR 142m. Close-out netting reduces this to EUR 38m, the single largest step, created by a legal document and a clean opinion. Variation margin reduces it to EUR 12m, the residual being the MPoR move plus threshold and MTA. Initial margin reduces it to approximately zero. Each step is a different kind of mitigation: netting is legal, VM is operational, IM is financial. Removing any one of them changes the answer by an order of magnitude, and each is a separate thing to get right. All figures illustrative only.

Margin call operations are where the theory meets a phone call

The daily sequence is unremarkable until it is not. The collateral system values the portfolio, computes the required collateral, compares it to holdings, applies threshold and MTA, and issues a call to the counterparty by an agreed deadline. The counterparty agrees and delivers, or disputes.

A dispute means the two parties value the same portfolio differently. The causes are mundane and persistent: a trade in one book and not the other, different market data, different models for a non-standard trade, a different treatment of a cashflow paid today. The resolution process is portfolio reconciliation — compare trade populations, then compare valuations trade by trade, isolate the drivers. Where the dispute cannot be resolved, the parties typically exchange the undisputed amount and escalate the remainder, and regulation requires material disputes above defined thresholds and ages to be reported to supervisors.

A failure to deliver collateral is a failure to pay, and a failure to pay is an event of default under the ISDA Master, subject to the grace period in the Schedule. This is the ordinary path to a default: not a bankruptcy filing but a margin call that goes unmet on a Tuesday. The escalation from a missed call to a default notice is therefore one of the most consequential operational workflows a bank runs, and it should be governed by a documented decision framework rather than improvised under pressure, because the decision to declare a default against a large counterparty is a decision to crystallise the loss.

The critical point — collateral is a transformation, not a deletion

A fully collateralised derivatives book has almost no credit exposure and a very large, very volatile, very unpredictable cash requirement. The risk has not gone away. It has changed shape, from a slow risk that materialises on default to a fast risk that materialises on a market move.

Think about what a bank or a client actually signs up to under a daily zero-threshold CSA. Every day, the mark-to-market of a portfolio worth billions of notional is settled in cash. On a quiet day the transfer is small. On a day when EUR/USD moves three per cent, or gas prices double, or gilt yields jump a hundred basis points, the transfer is enormous, it is due tomorrow morning, and it must be in cash. Nobody defaulted. Nothing about anybody's credit quality changed. The contract simply demanded cash.

This is why credit risk and liquidity risk are not alternatives; collateral trades one for the other. The trade is usually worth making, because a bank would far rather owe cash to a solvent counterparty than hold an unsecured claim on a failing one. But the liquidity risk it acquires is genuinely dangerous, for three reasons. It is procyclical, since calls are largest when markets are most stressed and funding is hardest to obtain. It is correlated, since every participant faces calls on the same day, driven by the same move, so everyone raises cash simultaneously by selling the same assets. And it is non-negotiable, since a margin call cannot be deferred, restructured or refinanced, and failing to meet one is a default.

The 2022 experience made all three visible at once. European utilities hedging their power and gas output faced tens of billions of euros of margin calls as prices spiked, and several governments extended emergency liquidity lines to firms that were economically hedged, commercially sound and simply out of cash. In the UK, the September 2022 gilt sell-off forced liability-driven investment funds to sell gilts to meet collateral calls on their interest-rate hedges, which pushed yields higher, which generated further calls — a self-reinforcing loop that ended only with Bank of England intervention. In neither case did a counterparty's credit quality drive the outcome. The collateral mechanism itself was the transmission channel. Part 5 develops both cases in full.

How smaller counterparties reach the market, and what it costs them

Not everyone can be a clearing member of a CCP or run a SIMM calculation. The access ladder determines what a smaller counterparty pays and in what form.

Direct clearing membership requires capital, operational capability and a contribution to the CCP's default fund. It is available to large dealers and a small number of large buy-side firms.

Client clearing is the standard route: a client faces a clearing broker, the broker faces the CCP, and the client posts margin to the broker. The client gets the CCP's risk profile and pays the broker a fee for access, balance sheet and default-fund contribution. The economics of this business have compressed sharply since 2008 as capital rules made client clearing expensive to provide, and the number of providers has shrunk.

Indirect clearing extends the chain one link further, letting a client of a client reach the CCP. It works, and it introduces exactly the layered-intermediary structure that makes the treatment of client collateral in a broker default the important question. Account segregation models, omnibus versus individually segregated, determine whether a client's margin is identifiable and portable to another broker when its own broker fails.

And then there is Helvetia, which does none of this. As an NFC− it is exempt from the clearing obligation and from the margin requirements. It posts nothing, holds no cash against the hedge, and its treasurer sleeps well, which is exactly what the exemption was designed to achieve: a corporate hedging a genuine commercial exposure should not be forced into the liquidity risk that hedging with margin creates.

But the risk did not disappear; it was priced. The bank charged EUR 4.2m of CVA plus EUR 1.1m of FVA, so EUR 5.3m embedded in the swap rate. Under a zero-threshold daily CSA the same trade would carry a CVA well under EUR 0.5m, and Helvetia would instead be exposed to margin calls that could plausibly reach EUR 40m or more in a stressed EUR/USD move. The trade-off is explicit: about EUR 5m paid up front and certain, against a contingent liquidity requirement of tens of millions that arrives when it is least convenient. For a corporate with a bond issue to hedge and no derivatives treasury operation, that is a rational choice, and it is the right way to explain the CVA charge to a client who complains about it. The client is not being taxed. The client is buying the liquidity insurance it declined to self-provide. Illustrative only.

Collateral management is the most process-intensive part of the counterparty-risk stack and the one with the clearest automation payoff, because the workflow is high-volume, rule-based and adversarial.

The core system holds the CSA parameters as structured data, values the portfolio daily, computes calls, issues and receives them over an electronic messaging channel, tracks agreement and settlement, books the collateral movements, and manages substitutions and interest. Alongside it sits a reconciliation engine comparing trade populations and valuations with counterparties, and a dispute workflow tracking cause, age and resolution.

Three architectural points decide whether it works at scale. The CSA parameters must come from the documentation system rather than being maintained locally. The valuation used for margin should be the same valuation the risk and finance functions use, or the bank will dispute internally before it disputes externally. And collateral inventory must be managed centrally across the whole firm, because the optimisation question — which asset to deliver against which call, given eligibility, haircuts and funding cost — cannot be answered by a system that only sees one counterparty.

The direction of travel is straight-through processing with exception-only human involvement, algorithmic collateral optimisation, and automated dispute resolution that identifies the driving trade rather than reporting a difference. Part 5 develops this.

  1. What does variation margin actually do to the exposure profile?

  2. Why does initial margin exist if variation margin is already being exchanged?

  3. Helvetia posts nothing. What is it paying instead?

Step 9 — Hedging and capital turn a reserve into a managed position

Having priced the risk and taken it onto the balance sheet, the bank now has to live with it for ten years. Two things follow: the XVA desk must stop the reserve from moving its P&L around, and the capital function must hold regulatory capital against the exposure. Both are continuous activities, and both are where the economics of the trade are finally settled.

Hedging CVA means hedging two different things at once

CVA moves for two reasons, and they require two different hedges. It moves when the counterparty's credit spread moves, because the default probabilities in the formula change. And it moves when the market moves, because the exposure profile in the formula changes. Both are live from the day the trade books.

Credit hedging is straightforward where a liquid single-name CDS exists. The desk buys protection on the counterparty in a notional sized to the CVA's sensitivity to that name's spread. When the spread widens, the CVA rises, which is a loss, and the CDS gains, offsetting it. The hedge is dynamic: as exposure and spread move, the required notional moves, so the position must be rebalanced.

Market-factor hedging addresses the exposure side. The CVA on the Helvetia swap is a function of EUR/USD and of EUR and USD interest rates, because those determine the exposure profile. Differentiate the CVA with respect to them and you get a set of deltas that can be hedged in the FX and rates markets. These hedges are not the same as the hedges on the underlying swap, and this is the point people find counter-intuitive. The rates desk hedges the swap's own market risk. The XVA desk separately hedges the market sensitivity of the credit charge on that swap, a second-order position that behaves like a portfolio of options, because exposure is the positive part of a value and options are what positive-part functions are.

The proxy-hedging problem is Helvetia's exact situation and it is the normal condition, not the exception. Helvetia has no traded CDS. There is no instrument the desk can buy that pays out on Helvetia's default. The desk's options are all imperfect. It can buy the iTraxx Europe index or a suitable index tranche, which hedges the systematic component of credit spread movement, the part of Helvetia's spread that moves because all European credit moves. It can buy protection on a liquid single name in the same sector and rating band, a European chemicals BBB issuer, which captures more of the sector-specific move. It can buy protection on a basket of chemicals names.

What none of these hedges is Helvetia's own idiosyncratic risk: a failed plant, a litigation loss, a leveraged acquisition, a fraud. That residual is called basis risk and it is unhedgeable by construction. The bank has hedged the part of the loss that comes from the market re-pricing all chemicals credit, and retained the part that comes from Helvetia specifically being in trouble, which is, of course, the part most likely to cause an actual default. The honest description of a proxy-hedged CVA position is that it is hedged against the scenario in which the bank probably survives and unhedged against the scenario in which it does not.

Two further constraints bite. Buying CDS on a client without the client knowing is commercially delicate and, in some markets, reputationally difficult. And the CVA hedge is itself a derivative position with a counterparty, usually another dealer, usually collateralised, which generates its own funding requirement. That is the FVA of EUR 1.1m in Helvetia's charge: the cost of posting variation margin on the collateralised hedge of an uncollateralised client trade.

Figure 9 What the XVA desk can hedge, can only proxy, and must simply reserve. Illustrative figures.

The decomposition explains why the desk's job is uncomfortable. The market components — the sensitivity of the CVA to EUR and USD rates and to EUR/USD spot and basis — are large and cleanly hedgeable in deep markets, and the desk deals with them in the ordinary course. The credit component is large and, for a name like Helvetia with no traded CDS, hedgeable only through a proxy: an index, a sector basket, or a liquid name judged to behave similarly. The proxy removes most of the systematic movement and leaves the idiosyncratic part untouched, so the desk is hedged against European industrial credit repricing and unhedged against Helvetia specifically having a bad year.

What remains cannot be hedged at all, and must simply be held. The cross-gamma between credit and FX, the recovery rate assumption, and the wrong-way correlation are not traded quantities. There is no instrument to buy. The desk reserves against them, the model validation function argues about the assumptions, and the CVA capital charge exists precisely because supervisors know this residual is there. A useful way to read Figure 9 is as a map of the desk's honesty: green is what it manages, amber is what it approximates, and red is what it is exposed to no matter how well it runs its book.

SA-CCR turns a derivative into a loan-equivalent number

Regulatory capital requires a single exposure number, so the framework converts a distribution into a loan equivalent. The Standardised Approach for Counterparty Credit Risk (SA-CCR) is the current standardised method, applying in the EU from June 2021 under CRR2. It replaced the Current Exposure Method and the Standardised Method, both of which were criticised for ignoring netting properly, ignoring collateral properly, and failing to distinguish margined from unmargined trades.

The formula is short: EAD = 1.4 × (RC + PFE add-on).

RC is the replacement cost, what it would cost to replace the netting set today. For an unmargined netting set it is the greater of zero and the net mark-to-market less net collateral held. At inception a swap is at fair value, so RC is zero.

The PFE add-on is a supervisory proxy for future exposure, built bottom-up. Trades are allocated to asset classes — interest rate, FX, credit, equity, commodity — and within each to hedging sets that recognise offsetting. Each trade gets an adjusted notional reflecting its size and duration, a supervisory delta reflecting direction and optionality, and a maturity factor. A prescribed supervisory factor per asset class scales the result. The asset-class add-ons are then aggregated with prescribed rules and multiplied by a multiplier that gives limited recognition to over-collateralisation and negative mark-to-market.

The alpha of 1.4 is a fixed supervisory multiplier applied to the whole. It carries over from the Internal Model Method, where it was calibrated to bridge the gap between an EEPE computed from a bank's own simulation and a loan-equivalent exposure appropriate for the capital formula. It is doing three jobs at once: adding conservatism for general wrong-way risk, adding conservatism for portfolio concentration, and covering the fact that the capital formula was designed for loans with a fixed exposure rather than derivatives with a stochastic one. It is not a modelling parameter and it cannot be argued down. It is a policy buffer, and its practical effect is that every derivative exposure is capitalised forty per cent above its computed level.

Figure 10 SA-CCR: turning a derivative into a loan-equivalent exposure. Illustrative figures.

The alpha of 1.4 is a supervisory thumb on the scale, and the framework is honest about it being one. It exists because a closed-form formula producing a single exposure number cannot capture correlation between exposure and default, portfolio granularity effects, or the general possibility that the formula is optimistic. Rather than model those things, the framework multiplies the answer by forty per cent and moves on — a blunt instrument, deliberately preferred to an elegant one that supervisors would then have to validate bank by bank.

IMM is the alternative, and fewer banks now think it is worth it

The Internal Model Method (IMM) lets a bank use its own simulation to compute EAD, as alpha multiplied by effective EPE, with alpha normally 1.4 and the effective EPE coming from the bank's own exposure engine. It requires supervisory approval, a demonstrated track record, independent validation, backtesting, stress testing, and a use test showing the model genuinely drives internal risk management rather than existing to produce a capital number.

IMM produces lower capital than SA-CCR for well-diversified, well-collateralised portfolios, because it recognises netting and collateral as they actually work rather than through prescribed factors. The reason fewer banks now pursue it is that the payoff has shrunk while the cost has not. Approval is slow and uncertain. Maintenance, validation and backtesting are permanent costs. And the Basel 3.1 output floor, as it is being applied across jurisdictions, caps how far internal models can reduce capital relative to the standardised approach, which truncates the benefit precisely where the model was most favourable. A bank that must in any case compute SA-CCR to determine its floor is running two frameworks to obtain a capped benefit on one. Many have concluded that the arithmetic no longer works, and have kept their simulation for pricing, limits and CVA while using SA-CCR for capital.

The CVA capital charge is separate, and it exists because of what happened in 2008

Roughly two-thirds of counterparty-credit losses in 2007 to 2009 came from CVA mark-to-market moves rather than from actual defaults. Banks held capital against the risk that counterparties would default, and lost money instead on the revaluation of the credit charge as spreads widened on counterparties that never defaulted at all. Basel III introduced a dedicated CVA capital charge in 2010 to close that gap.

The framework was subsequently revised as FRTB-CVA, offering two approaches. BA-CVA, the basic approach, is formula-driven, available in a reduced version that ignores hedges and a full version that recognises eligible single-name and index credit hedges. It is intended for banks with smaller or simpler CVA books. SA-CVA, the standardised approach, is sensitivity-based and structurally similar to the market-risk standardised approach: the bank computes the CVA's sensitivities to credit spreads, interest rates, FX and equity, applies prescribed risk weights and correlations, and aggregates. It requires supervisory approval and the ability to compute CVA sensitivities accurately, which in turn requires the AAD machinery described in Step 7. It generally produces lower capital for a hedged book, because it recognises a wider range of hedges.

The design point that matters is that both approaches recognise hedges, and the recognition is what makes the XVA desk's hedging economically rational. A desk that hedges its CVA and receives no capital credit for doing so is spending money to reduce accounting volatility while its capital charge stays flat. Under FRTB-CVA the hedge reduces both.

Compression removes trades rather than risk, and it is the only lever here that cuts capital without cutting business

Every capital and operational charge in this module is levied on the population of trades, not on the net risk position the bank actually holds. SA-CCR computes an add-on from adjusted notionals. The leverage ratio counts exposures gross. Confirmation, reconciliation, margin processing, reporting and the nightly simulation all scale with the number of line items. A dealer that has traded with the same four counterparties for fifteen years accumulates thousands of contracts that partially offset each other, and it pays for all of them while carrying the risk of almost none of them. Compression is the exercise that fixes the mismatch.

Compression is a multilateral tear-up. Participants submit their portfolios and their tolerance for residual risk change to a compression service. The service searches for sets of trades across all participants that can be terminated, together with a small number of replacement trades, such that every participant's net risk position stays inside its stated tolerance. Participants accept or reject the proposed cycle, and accepted terminations are legally novated so that the old contracts genuinely cease to exist rather than merely being netted in a report. The economics are unusual: nobody gives anything up, and everyone's capital and operational cost falls, because the thing being removed was never risk in the first place.

The distinction between compression and rebalancing matters. Compression terminates redundant trades and leaves the risk profile as close to unchanged as the tolerance permits. Rebalancing deliberately moves risk between participants — most often to move positions towards a clearing house or away from a constrained balance sheet — and it is a different exercise with different approvals. Confusing the two is how a compression cycle turns into an unauthorised position change.

The scale of this is easy to underestimate. Compression is the principal reason global OTC derivatives gross notional outstanding fell from its 2013 peak even as trading activity continued to grow. It is one of the few genuinely large structural changes in the post-crisis market that happened without a rule requiring it, because the capital framework made it economically obvious. Third-party post-trade infrastructure providers run compression cycles for bilateral books, and clearing houses run the equivalent for cleared portfolios, where the multilateral nature of the problem is even more favourable.

Figure 11 Compression removes trades, not risk — and SA-CCR is charged on the trades. Illustrative figures.

KVA closes the loop

Capital has a cost, and a trade that consumes capital for ten years must pay for it. Helvetia's trade consumes EUR 2.7m of capital at inception under SA-CCR, plus the CVA capital charge, declining over the life as the add-on's maturity factor falls. If the bank's shareholders require, say, a ten per cent return on equity, then holding roughly EUR 2.7m for a decade requires the trade to generate a present value of returns that the market price of the swap does not automatically provide.

KVA is that cost, computed at inception and charged in the price. For Helvetia it is EUR 0.9m over the life. It completes the total XVA charge of EUR 6.2m from Step 5: CVA EUR 4.2m for the expected credit loss, FVA EUR 1.1m for the funding of the hedge, KVA EUR 0.9m for the capital consumed.

This is where the module's through-line becomes concrete. The bank has now measured the risk as an exposure in Step 7, priced it as a derivative in Step 5, funded it as a balance-sheet item through FVA and MVA, and hedged it as a market position in Step 9. Those four things are the whole job. Helvetia's trade has had all four done to it, and the total price of doing them is EUR 6.2m on a EUR 460m notional, which is 135bp, or about 17bp per annum. That is what it costs to take ten years of uncollateralised credit risk on a BBB− corporate and manage it properly.

Capital calculation is a reporting problem that behaves like a risk problem, and the difficulty is reconciliation rather than mathematics.

SA-CCR is arithmetically simple but data-hungry: every trade needs an asset class, a hedging set, an adjusted notional, a supervisory delta and a maturity factor, all derived from attributes set at booking in Step 6. Errors here are systematic rather than random, because a mis-classified product family mis-capitalises every trade in it.

The recurring architectural failure is that the capital calculator, the risk engine and the finance ledger each build their own view of the trade population, and the three do not agree. The bank then cannot explain why its RWA moved, which is the question the regulator, the CFO and the board all ask. The fix is a single reconciled trade population feeding all three, with the differences between measures explained by design rather than discovered by investigation.

SA-CVA additionally requires accurate CVA sensitivities across thousands of risk factors, which is not feasible by finite difference at portfolio scale. AAD is effectively a prerequisite. This is one of the few places where a quantitative technique is directly a regulatory-approval dependency.

Stress testing asks the question the daily number cannot: what if history is not the guide

The nightly exposure calculation is calibrated on what markets have done, and the events that destroy banks are the ones markets have not yet done. That is not a criticism of the model; it is a statement of what a model is. A simulation calibrated on the last five years of EUR/USD will produce a faithful distribution of the next five years conditional on the world continuing to behave as it has. Stress testing exists to break that condition deliberately, and it is a separate discipline from the daily cycle rather than an extension of it.

Stressed calibration is the regulator's minimum, and it is not the interesting part

Supervisors require exposure models to be calibrated to a period of stress as well as to recent history, producing a stressed effective expected positive exposure alongside the ordinary one. The capital number takes the more conservative of the two. The mechanism is straightforward — re-estimate volatilities and correlations from a chosen stress window, typically one containing the 2008 crisis, and re-run. The value is real but bounded: it protects against a bank quietly benefiting from a calm calibration period, which was a genuine problem before the rule existed. It does not answer any question the bank did not already know to ask.

The hard part is stressing exposure and credit quality together

Stressing the market and stressing the counterparty separately produces two numbers that understate the problem when combined. The interesting scenario is not "EUR/USD moves twenty per cent" and it is not "Helvetia is downgraded three notches". It is both, in the same scenario, because they are connected — and this is where wrong-way risk stops being a concept and becomes a number the board can look at. Building this properly means the credit model and the market model have to share a scenario, which sounds obvious and is architecturally awkward in most banks, because the two were built by different teams against different rules in different decades.

Scenario design has three families and each answers a different question. Historical replay takes an actual episode and runs the current book through it, which is credible precisely because it happened and which is limited to the shocks history has supplied. Hypothetical scenarios are constructed to target a known vulnerability, which makes them sharper and makes their plausibility a matter of judgement. Regulator-prescribed scenarios arrive annually, apply across the industry, and answer a comparability question rather than a management one — they tell a supervisor how banks rank against a common shock, not what any individual bank should worry about.

Four questions a board should be able to have answered within a day

The measure of a counterparty stress framework is not its sophistication but whether it answers plain questions quickly. Which single counterparty costs us most if it defaults tomorrow, and is that name on any watchlist. Which market move produces the largest aggregate margin call across the book, and can we fund it. What is our largest plausible single-day collateral outflow. And which counterparties would breach their limits under the scenario we are most worried about, before they breach them in reality. A bank that needs three weeks to answer these has a data problem rather than a modelling one, which is the point Part 4 develops.

Reverse stress testing starts from the failure and works backwards

Ordinary stress testing asks what a scenario does to the bank; reverse stress testing asks what scenario would break it. The starting point is a defined failure — the loss that exhausts a capital buffer, the collateral call that cannot be funded, the concentration that becomes fatal — and the exercise is to construct the conditions that produce it and then judge how far-fetched they really are. It is uncomfortable by design, because the honest output is often a scenario that is neither extreme nor unlikely.

This is the discipline that would have caught Archegos. No ordinary scenario asks "what if one client's positions are ten times what we think and eight other banks hold the same trade". A reverse exercise starting from "we lose five billion on a single prime brokerage client" arrives at exactly that question, because it is the only path to a loss of that size. Part 5 develops the case; the point here is that the tool existed and the question was not asked.

Stress testing is a compute problem of a different shape from the nightly cycle, and treating it as the same problem is a common architectural mistake. The nightly run is narrow and deep under hard time pressure: one scenario set, full trade population, twelve hours. Stress is broad and shallow with no overnight deadline: hundreds of scenarios, full revaluation, and an appetite that expands to whatever capacity exists. The nightly grid is sized for latency and the stress estate is sized for throughput, which is why stress workloads are the first thing most banks moved to elastic cloud capacity — the demand is spiky, the deadline is soft, and nobody has to explain a missed batch at six in the morning.

Step 10 — The loss happens at close-out, but it was decided at documentation

Everything in the previous nine stages was preparation for this one, and by the time it arrives almost nothing can still be changed. The netting works or it does not. The collateral is held or it is not. The documentation is clean or it is contested. Stage 10 is where the bank discovers what its earlier decisions were worth.

Distress is visible long before default, if anyone is looking

Most defaults are preceded by months of signals, and the value of a watchlist process is that it forces someone to act on them before the options run out. The signals fall into four groups.

Rating migration. A downgrade, or a negative outlook, or the internal rating model producing a worse grade on updated financials. Migration to sub-investment grade is a step change, because it removes the counterparty from many investors' eligible universes and raises its funding cost sharply.

Spread widening. Where a CDS or a liquid bond exists, the market prices deterioration continuously and does so faster than any rating agency. Where it does not, which is Helvetia's case, the proxy curve moves with the sector and misses the idiosyncratic story entirely, which is exactly when a credit officer's judgement has to substitute for a market price.

Fundamental and covenant signals. A breach or near-breach of a financial covenant on the counterparty's bank debt. A drawn revolving credit facility that was previously undrawn. An auditor's going-concern qualification. A resignation of the CFO. A delayed filing.

Market and behavioural signals. Equity price falls and rising equity volatility. Bond prices trading at distressed levels. And the operational signals from Stage 8: late collateral deliveries, an increase in disputes, requests to amend terms, a sudden interest in unwinding trades at unusual levels.

Watchlist status triggers a defined response: no new business without senior approval, a reduction plan for existing exposure with a timetable, more frequent review, and often a shift of the relationship from the business to a specialist workout team.

Mitigation before default is a race against the counterparty's remaining optionality

Once a counterparty is in trouble, every mitigation is available at a worse price than it was a month earlier, and some are no longer available at all. The choices are these.

Restructure the trade. Shorten the maturity, reduce the notional, or add amortisation, in exchange for a payment. This works while the counterparty still has something the bank wants and vice versa.

Novate to another dealer. Transfer the trade to a third party who accepts the counterparty risk in exchange for a payment. The bank exits entirely. The constraint is that another dealer has to want the position, which becomes harder exactly as it becomes more urgent, and the counterparty must consent.

Unwind at mid. Terminate by agreement at a fair value. Clean, and it requires the counterparty to have the cash to pay if it is out of the money, which a distressed counterparty typically does not.

Buy protection. Where a CDS exists, buy it on the specific name. The cost has already risen with the spread, so this converts a contingent loss into a certain cost. It is a decision to pay now to avoid paying more later, and it is usually taken too late.

Exercise a mandatory break clause. A mandatory break is a contractual right, negotiated at inception, to terminate a long-dated trade at fair value at a stated intermediate date, typically at year five of a ten-year trade. It converts a ten-year credit exposure into a five-year one with an option to re-engage, and it caps the CVA accordingly.

Banks began demanding break clauses systematically after 2008 for a simple reason: they discovered they had no way out of long-dated uncollateralised positions with deteriorating counterparties. A break clause is the exit that a CSA-free relationship otherwise lacks. Helvetia's Schedule does not contain one, which was a commercial concession at inception and becomes a constraint in Step 10. This is the clearest illustration in the module of the principle that the close-out is decided at documentation: the bank's options in year three were fixed in week eleven of Stage 2.

Negotiate a CSA into the existing relationship. Possible, and the price of asking is that the request itself signals that the bank has changed its view of the client, which can accelerate the deterioration it is trying to protect against.

Events of Default and Termination Events are the contractual triggers

The ISDA Master distinguishes Events of Default, which involve fault, from Termination Events, which do not. The distinction matters for who gets to terminate and how the amount is determined.

The Events of Default include failure to pay or deliver, subject to a short grace period, which is the one that actually fires, usually as an unmet margin call. Bankruptcy, broadly defined to cover insolvency proceedings, administration, liquidation and analogous processes across jurisdictions. Cross-default, where a default on the party's other borrowed money above a stated threshold triggers a default under the ISDA even though every payment under the ISDA has been made. Breach of agreement, misrepresentation, and default under specified transaction.

Termination Events include illegality, tax events, credit event upon merger, and the Additional Termination Events negotiated in the Schedule, being the downgrade triggers, NAV triggers and covenant triggers described in Stage 2.

Cross-default deserves attention because it is the clause that turns a bilateral problem into a market-wide one. A counterparty that defaults on a syndicated loan has, by virtue of cross-default, simultaneously defaulted under every ISDA it has signed. All its derivative counterparties acquire termination rights on the same day. The synchronised close-out that follows is a feature of the documentation, not of the counterparty's cash position, and it is why the threshold in the Schedule is a genuinely consequential negotiation.

The close-out mechanics

The non-defaulting party has a right, not an obligation, and the first decision is whether to exercise it. Terminating crystallises the loss and ends any chance of the position recovering. Not terminating leaves the bank exposed to a defaulted counterparty. In practice, where the counterparty has entered formal insolvency the decision is straightforward; where it has merely missed a payment, it is a judgement.

If the bank terminates, it designates an Early Termination Date, giving notice as the Master requires. All transactions terminate on that date. Each is then valued, and the valuation methodology depends on which version of the Master governs.

Under the 1992 Master, the parties elected between Market Quotation, meaning obtain quotations from reference market-makers for replacement transactions and take the average, and Loss, a broader good-faith determination of total losses and costs. Market Quotation was intended to be objective and turned out to be unworkable in a genuine crisis, because in September 2008 there were no market-makers willing to quote replacement transactions on a portfolio of thousands of trades from a defaulting counterparty. The objective standard failed exactly when it was needed.

Under the 2002 Master, both were replaced by the Close-out Amount: the amount of losses or costs, or gains, that the determining party incurs in replacing the material terms and option rights of the terminated transactions, determined in good faith using commercially reasonable procedures to produce a commercially reasonable result. The determining party may use quotations, market data, and its own internal models, explicitly including internal models where external data is unavailable. It is a flexible standard designed to work in a dislocated market, and its flexibility is precisely what makes it litigable.

The values are then converted into the termination currency and summed into a single net amount. The Master's set-off provision may permit further set-off against amounts owing under other agreements between the same parties. The result is one figure payable in one direction. If it is owed to the bank, the bank files an unsecured claim in the insolvency estate, applies any collateral it holds, and waits, often for years, for a distribution.

Close-out is where valuation becomes litigation

The structural problem is simple to state and impossible to solve: the surviving party values its own claim. The bank determines the Close-out Amount, and the bank is the party that will be paid it. It is required to act in good faith and to use commercially reasonable procedures, and there is no independent valuer, no exchange price and no arbiter until a court is asked.

Lehman Brothers is the case that established what this means at scale. The failure required the close-out of roughly 900,000 derivative transactions across 6,500 legal entities. Litigation over close-out valuations ran for more than a decade, and the case law it generated — on what commercially reasonable means, on whether a determining party may include its own funding and credit costs, on the treatment of bid-offer and liquidity adjustments in a dislocated market, on the validity of provisions that suspend payment obligations to a defaulted party — is now the practical content of the 2002 Master's flexible standard.

Three lessons carry forward. Close-out netting worked where documentation was clean and failed where it was not, so the mechanism is sound and the execution is where the losses were. The dispute is about the middle of the process, not the principle, since nobody argued about whether netting applied, they argued about what a portfolio was worth on a Tuesday in September when there was no market. And the operational burden was the binding constraint, because identifying which trades sat under which agreement with which entity, and locating the collateral, took years, and that work is done in the systems described in Stage 6, not in a courtroom.

The CCP alternative replaces an exposure with a place in a queue

Central clearing changes the counterparty, and it changes the nature of what happens on default. Under novation, the original bilateral contract between the bank and its counterparty is extinguished and replaced by two contracts: one between the bank and the central counterparty (CCP), and one between the CCP and the original counterparty. The CCP becomes buyer to every seller and seller to every buyer. The bank's credit exposure is now to the CCP.

The CCP protects itself with a default waterfall, an ordered sequence of financial resources consumed in a defined order when a clearing member fails.

Figure 12 The CCP default waterfall: clearing replaces an exposure with a place in a queue.

The defaulter's initial margin goes first. Collateral posted by the failed member, sized to cover the expected cost of closing out its positions over the margin period of risk. In most historical defaults this layer alone has been sufficient.

The defaulter's default fund contribution is next. Every clearing member contributes to a mutualised fund, and the defaulter's own contribution is consumed before anyone else's.

The CCP's own capital, the skin in the game, comes third. A tranche of the CCP's equity, deliberately placed ahead of surviving members' contributions so that the CCP has a direct financial incentive to margin its members adequately. It is typically small relative to the fund, and how small is a live policy debate.

Surviving members' default fund contributions are fourth. This is the mutualisation layer: losses from one member's failure are now borne by the others.

Assessment rights, being powers to call additional contributions from surviving members up to a capped multiple of their existing contributions, are fifth.

Recovery tools are last: variation margin gains haircutting, which writes down the payments owed to members with profitable positions, partial tear-up of contracts, and ultimately resolution of the CCP itself.

The argument to take from this is that clearing mutualises risk; it does not eliminate it. Bilaterally, the bank faces the risk that its counterparty fails. Centrally, the bank faces the risk that another clearing member fails badly enough to burn through the first three layers and reach its own default fund contribution — a member it did not choose, cannot assess and has no relationship with. It has swapped a known idiosyncratic risk for an unknown mutualised one, and it has concentrated a large part of the market's counterparty risk into a small number of institutions whose failure would be a systemic event. That trade is generally worth making, since CCPs are heavily regulated, margin their members daily, run tested default-management procedures, and have an excellent track record. But it is a trade, and describing clearing as removing counterparty risk is wrong in a way that matters for how a bank measures and capitalises its CCP exposures. Part 3 develops the CCP as an actor in the market structure.

The default management process is an auction, and the incentives inside it are deliberate

When a clearing member defaults, the clearing house does not simply hold the portfolio and hope. It first hedges the defaulted book to stabilise it, using the default fund resources available, and then auctions the portfolio — usually in currency or product lots — to the surviving members. The auction is the mechanism that converts an unmanageable position into someone's ordinary trading book, and it has to work quickly, because every day the portfolio remains unhedged is a day of losses eating through the waterfall.

Surviving members are required to participate, and bidding badly is penalised. Members who submit uncompetitive bids can be juniorised: their default fund contributions are consumed earlier in the waterfall than those of members who bid well. The design is a direct answer to an obvious problem — a member who expects to absorb losses through mutualisation has every incentive to bid low and buy the book cheaply, which is exactly the behaviour that would make the auction fail. Juniorisation prices that temptation.

Recovery and resolution ask who pays when the waterfall runs out

The waterfall is finite, and the question of what happens after it is exhausted is the live policy debate in this part of the market. Once member contributions and assessment rights are consumed, a clearing house moves into recovery: tools written into its rulebook that allocate remaining losses to participants. Variation margin gains haircutting reduces payments to the members whose positions gained, which is the last layer of the waterfall in Figure 12 and the first genuinely contentious one, because it takes money from firms that did nothing wrong. Forced allocation assigns portions of the defaulted portfolio to members who did not bid for them. Full tear-up terminates the affected contracts entirely at a determined price, which ends the clearing house's exposure by ending the market.

If recovery fails, resolution follows, and resolution is a statutory process rather than a contractual one. A resolution authority takes control, and the tools available resemble those used for banks: write-down of claims, transfer of positions to a bridge entity, and allocation of losses according to a statutory hierarchy rather than the rulebook. Several jurisdictions have built dedicated CCP recovery and resolution regimes precisely because the alternative — a disorderly failure of an institution to which the entire market has been directed by mandate — is unacceptable.

The honest summary is uncomfortable and worth stating plainly. Policy after 2008 concentrated counterparty risk into a small number of clearing houses on the judgement that a small number of heavily regulated, heavily margined, transparently governed nodes is safer than a dense web of bilateral exposures. That judgement is probably right. It is also a decision to create single points of failure, and the recovery and resolution framework is the admission that the tail was not eliminated but relocated.

Recovery, realised loss, and what the CVA was actually for

When the dust settles, the bank has a net claim, some collateral, and a recovery rate. If the Close-out Amount owed to the bank is EUR 40m, it holds no collateral, and the estate ultimately distributes forty cents on the euro after three years, the bank recovers EUR 16m and realises a loss of EUR 24m.

Compare that to the CVA charged at inception. The CVA was EUR 4.2m. The realised loss, in this scenario, is EUR 24m. Was the CVA wrong?

No, and understanding why is the last conceptual step in this part. The CVA was a price for a distribution, not a prediction of an outcome. It was the probability-weighted expected loss across every future state of the world, including the roughly sixty-nine per cent of them in which Helvetia never defaults and the loss is zero, and the states in which it defaults early when exposure is small, and the states in which it defaults late when exposure is large. The realised outcome is one draw from that distribution. A single draw cannot validate or invalidate a price, any more than one house burning down proves that the insurer mis-priced the premium.

What can validate the price is the aggregate. Across hundreds of counterparties and thousands of trades over many years, the CVA charged should approximate the losses realised plus the cost of the hedges bought. If it consistently does not, the bank is mis-pricing. That comparison, cumulative CVA charged against cumulative realised loss and hedge cost, is the only meaningful backtest of an XVA franchise, and remarkably few institutions produce it.

The moment that matters most — Helvetia at year 3, with no default at all

Now the teaching point that the whole module is built around.

It is March 2029. The swap has three years behind it and seven to run. Helvetia has paid every coupon on time. It has not missed a payment, breached a covenant or filed for anything. By any ordinary credit standard it is a performing counterparty.

Two things have happened. EUR/USD has moved from 1.0870 to 0.94, a sharp dollar appreciation. The bank receives USD and pays EUR on the swap, so the position has moved substantially in the bank's favour: the swap is now worth +EUR 41m to the bank. And Helvetia has been downgraded to BB+, having lost a major contract and taken on debt for an acquisition that has not performed. Its proxy credit spread has widened from 220bp to 480bp.

Both moves push the CVA the same way. The exposure is larger, because the swap is deeply in the money and the remaining seven years of diffusion start from a much higher base. The default probability is higher, because the spread has more than doubled. Multiply a bigger exposure by a bigger probability of default and the answer is much bigger.

The CVA reserve on this one trade rises from EUR 4.2m to EUR 9.6m.

That EUR 5.4m increase is a realised profit-and-loss loss, booked today. Not a provision. Not a note in a risk report. Under IFRS 13 the CVA is part of the fair value of the derivative, so an increase in the CVA is a reduction in the fair value of the bank's asset, and it goes through the income statement in the period it occurs. The XVA desk reports a EUR 5.4m loss on a trade with a client who has paid every euro it owes.

Sit with what has actually happened. The bank has made a large gain on the swap itself, EUR 41m of positive mark-to-market. It has lost EUR 5.4m of that gain to the credit charge on the very position that generated it. The more valuable the trade becomes, the more it costs to hold, because value owed by a counterparty is credit exposure to that counterparty. This is the sense in which counterparty risk is genuinely different from every other risk on a bank's balance sheet: winning creates the exposure.

And the hedge only partly helps. The desk's iTraxx and sector proxy hedges gained on the general widening of European chemicals credit, but Helvetia's widening was substantially idiosyncratic, a lost contract and a bad acquisition rather than a sector move. Perhaps two of the EUR 5.4m is recovered from the hedges. The rest is the basis risk that Step 9 said was unhedgeable, arriving exactly as described.

Three conclusions follow, and they are the ones to remember from this part.

A counterparty-risk loss does not require a default. Roughly two-thirds of counterparty-credit losses in the 2007 to 2009 crisis came from CVA mark-to-market moves, not from defaults. This is that phenomenon, in miniature, on one trade.

The mitigation options available in year three were determined in Stage 2. With a break clause, the bank could terminate at year five. With a CSA, it would be holding EUR 41m of collateral and the CVA would be a fraction of EUR 9.6m. With a downgrade trigger at BBB− rather than BB−, it would have had a termination right. It negotiated none of those, at a moment when the client was strong, the competition was keen and the concession looked cheap.

And the EUR 4.2m charged at inception was not wrong. It was the correct price for the distribution as it stood in 2026. The distribution has now been re-drawn on worse information, and the difference between the two prices is the loss. That is what it means to say counterparty risk is a derivative: it is marked to market, every day, for ten years, whether or not anyone defaults. Illustrative only.

  1. What is the difference between EE and PFE, and why does a bank need both?

  2. Why is CVA a price rather than a fee?

  3. What does close-out netting protect against, specifically?

  4. Why does a CSA not eliminate exposure?

  5. What is the margin period of risk, and what are the standard assumptions?

  6. What is the difference between general and specific wrong-way risk?

  7. What is the alpha of 1.4 in SA-CCR actually doing?

  8. Helvetia's swap gains EUR 41m of value and the bank books a EUR 5.4m loss. Explain.

2

  1. Why set limits on potential future exposure rather than current MtM?

  2. Helvetia's group limit is 67 % used, its entity limit 91 %. Which governs?

  3. What does a concentration limit catch that counterparty limits miss?

3Counterparty risk is the only number in the bank that four functions must agree on

A market-risk number belongs to a trading desk. A credit-rating number belongs to a credit officer. A counterparty-risk number belongs to nobody and to everybody: the salesperson quotes off it, the trader hedges it, the credit officer limits it, the collateral team collects against it, Finance books it as a reserve, and the supervisor tests it. This part is about who those people are, what they owe each other, and why the arrangement is deliberately uncomfortable.

The four camps and why the bank is the only one that can see the whole picture

Every counterparty-risk decision is made at the intersection of four groups who want different things and see different fragments of the same exposure. They are not adversaries in any dramatic sense. They are simply optimising different objective functions with different information sets, and the shape of the business — the documents, the limits, the charges, the disputes — is the residue of their negotiation. Understanding who they are is not organisational trivia. It is the fastest route to understanding why the process in Part 2 has the shape it has.

The first camp is the counterparty, and its position is entirely rational: it wants the hedge and does not want to pay for its own credit. Helvetia Chemicals AG has issued a USD bond and carries a currency mismatch it did not choose and cannot manage. The cross-currency swap fixes a real problem. What Helvetia does not accept easily is that the swap comes with a credit price attached — the EUR 4.2m CVA charge in the running example — because from where the treasurer sits, the company is not borrowing anything. No money changes hands at inception. The bank is not extending a loan. And yet the bank is charging as if it had. That perception gap is the single most common source of friction in the pre-deal conversation, and Part 2 Stage 5 exists to close it.

The second camp is the bank itself, which is simultaneously salesperson, risk taker and risk manager — three roles that do not naturally coexist. The coverage banker is compensated on revenue and relationship. The trading desk is compensated on the market risk it manages well. The XVA desk is compensated on the counterparty risk it manages well. Credit risk is compensated on none of it, deliberately. The bank does not resolve this tension by pretending it does not exist. It resolves it by fragmenting authority — the salesperson may not agree a price the XVA desk has not quoted, the trader may not book a trade the limit system has not cleared, and the credit officer may not be overruled below a defined seniority. The organisation chart is a control.

The third camp is market infrastructure, and it exists because bilateral risk between two private parties turned out to be a public problem. Central counterparties (CCPs) interpose themselves between buyer and seller so that neither faces the other. Custodians and tri-party agents hold and move collateral so that neither party has to trust the other to segregate it. Trade repositories collect every derivative transaction so that a supervisor can, in principle, see the aggregate that no individual bank can. Infrastructure is not neutral plumbing. Each of these entities has its own rulebook, its own margin model, its own default waterfall and its own commercial interest, and a bank's exposure to a CCP is a real exposure that must be measured.

The fourth camp is the supervisor and the rating world, which sets the constraints inside which the other three negotiate. Regulators define what counts as an exposure (SA-CCR, the IMM), what capital must be held against it, what must be cleared and what must be margined, and what must be reported and to whom. Rating agencies and data vendors supply the credit inputs — the ratings, the CDS curves, the sector proxy curves — that let the bank convert a name into a hazard rate. This camp does not participate in any individual trade. It shapes every one of them.

The bank sits in the middle because it is the only party that can see the whole netting set

The bank's structural position is not a matter of prestige — it is a matter of information, and the information asymmetry runs one way. Consider what each party can observe. The counterparty sees its own trades with this bank and its own trades with every other bank, but it has no view of the bank's models, its funding cost or its capital consumption, and — more to the point — it has no incentive to volunteer the aggregate. The CCP sees only the cleared portion. The custodian sees only the collateral it holds. The trade repository sees everything but at a reporting lag, in a form nobody prices off, and with no authority to act. The supervisor sees the reported aggregate months after it mattered.

Only the bank sees every trade in the netting set at the moment it needs to decide. This matters because of the third of the six ideas in this module: the netting set, not the trade, is the unit of risk. A cross-currency swap that adds EUR 68m of peak potential future exposure standing alone might add almost nothing to a netting set where it offsets existing positions — or it might add more than its standalone figure if it concentrates the same risk factor. Nobody outside the bank can compute that number, because nobody outside the bank holds the full trade population, the legal netting opinion and the collateral terms in one place. Pricing authority follows information, and the information sits with the bank. That is why the charge is the bank's to calculate and the bank's to defend.

And it is precisely the failure of this claim at the industry level that Archegos exposed. Each of Archegos Capital Management's prime brokers could see its own netting set perfectly well. Credit Suisse could compute its exposure to Archegos. Nomura could compute its exposure to Archegos. Morgan Stanley could compute its exposure to Archegos. Not one of them could see that the same underlying equity positions had been replicated across several dealers on total return swaps, so that the true concentration in a handful of names was a multiple of anything any single book showed. The individual netting-set view was correct and the aggregate picture was invisible. Each bank held a locally accurate number and a globally meaningless one, and when the positions unwound the losses were roughly USD 5.5bn at Credit Suisse, roughly USD 2.9bn at Nomura and roughly USD 900m at Morgan Stanley. The lesson is not that the models were wrong. The lesson is that "the only party that can see the whole netting set" is a statement about one bank's own book, not about the counterparty's total position — and the gap between those two things is where the largest counterparty losses of the modern era have lived.

What actually flows between the camps, and when

The relationships are easier to hold in the head as flows attached to stages than as an organisation chart. The table below reads against the ten stages of Part 2 by name and number. Each row is a real, dated, auditable transfer of something — a document, a number, an instruction, a payment.

From To What flows At which stage
Counterparty Bank (coverage, credit) Financials, ownership structure, LEI, hedging rationale 1 — Onboarding and credit assessment
Bank (legal) Counterparty (legal) ISDA Master, Schedule, CSA drafts; negotiated close-out and threshold terms 2 — Legal architecture
External counsel Bank (credit, capital) Netting opinion by jurisdiction, permitting exposure to be measured net 2 — Legal architecture
Credit risk Trading and sales systems Approved limit grid: PFE, tenor, product, settlement, by legal entity 3 — The limit framework
Sales desk Limit system Pre-deal incremental exposure query on a hypothetical trade 4 — The pre-deal check
XVA desk Sales desk CVA, FVA, KVA components and the total charge to be embedded in the price 5 — XVA pricing and the charge
Sales desk Counterparty The all-in quoted price, credit charge included but rarely itemised 5 — XVA pricing and the charge
Trading desk Trade repository (via reporting) UTI, UPI, LEI pair, product and lifecycle data within the reporting window 6 — Booking and trade capture
Market data vendors Risk engine Curves, volatilities, FX fixings, credit spreads and proxy sector curves 7 — The daily exposure cycle
Risk engine Limit, collateral, capital and reporting systems EE, EPE, PFE profiles and EAD per netting set 7 — The daily exposure cycle
Collateral operations Counterparty collateral team Margin call, valuation statement, dispute notice 8 — Collateral and margin
Tri-party agent / custodian Both parties Collateral eligibility test, haircut application, segregation confirmation 8 — Collateral and margin
CCP Clearing member (the bank) Initial and variation margin calls, default fund contribution requirements 8 — Collateral and margin
XVA desk Credit hedge market CDS or index protection, FX and rates hedges of the CVA sensitivity 9 — XVA hedging and capital
Finance Auditors and supervisors The CVA reserve as an IFRS 13 fair-value component, with model evidence 9 — XVA hedging and capital
Bank (legal, workout) Counterparty, courts, administrators Early Termination notice, close-out amount calculation, claim filing 10 — Distress, default and close-out
Figure 13 The counterparty-risk actor map: the bank is the only party that sees the whole exposure.

The counterparty taxonomy: who you face determines what the risk is made of

Two counterparties with the same credit rating can generate counterparty risk that differs by an order of magnitude in size, shape and cost. This is the point at which counterparty risk separates decisively from lending. A loan to a BBB− corporate and a loan to a BBB− insurer are, to a first approximation, the same risk. A ten-year uncollateralised cross-currency swap with a BBB− corporate and a daily-margined equity swap with a BBB− hedge fund are not remotely the same risk, and no single-dimension credit view will tell you why. What follows is the taxonomy the business actually uses. Read it as the answer to a single question: for each type, what is the dominant thing that can go wrong?

Corporates: the uncollateralised long-dated case, and the reason the whole discipline exists

Corporates hedge real exposures, and that is both their virtue and the source of the problem. A non-financial counterparty enters a derivative because something in its business — a foreign-currency bond, a floating-rate loan, a commodity input, an export receivable — creates a risk it did not choose. Helvetia Chemicals AG is the archetype. It has issued USD 500m of ten-year debt and wants EUR liabilities, so it swaps. The hedge is not a position. It is the removal of a position.

The characteristics that follow from this are consistent and awkward. Corporate hedges are long-dated, because the underlying exposure is long-dated — a ten-year bond needs a ten-year swap, not a one-year one rolled ten times. They are usually uncollateralised, because a corporate treasury does not run a collateral operation, does not want to post cash against a hedge of an exposure it is not marking to market, and would find daily variation margin an intolerable drain on working capital. Under EMIR, a non-financial counterparty below the clearing threshold (NFC−) is exempt from both the clearing obligation and the margin requirements, so the exemption is not a loophole the bank exploits — it is the regime the regulator designed, precisely to avoid pushing liquidity strain onto the real economy. And most corporates have no traded CDS, so the credit input has to be built from a proxy: for Helvetia, the European chemicals BBB sector curve, 180bp at five years and 220bp at ten.

The dangerous version is wrong-way risk, and with corporates it is often structural rather than incidental. Wrong-way risk is the situation where the bank's exposure to a counterparty rises at exactly the moment the counterparty's credit quality deteriorates. It is easy to construct for a corporate: an airline hedging jet fuel is exposed to the fuel price, and if fuel collapses, the airline's hedge is deeply in the bank's favour at the same moment the airline's competitors are enjoying a windfall — or, in the reverse configuration, a commodity producer hedging its own output owes the bank most when the commodity that funds its cash flows has fallen. The general rule is to ask whether the derivative's payoff to the bank is correlated with the counterparty's business distress. When the hedge is on the commodity that drives the counterparty's revenue, the answer is almost always yes.

So why does the bank accept an uncollateralised ten-year exposure to a BBB− name with no traded CDS? Because the derivative is not the relationship. It is one product inside a relationship that also includes the revolving credit facility, the term loan, the bond issue the swap is hedging, the DCM fee earned on that issue, cash management, trade finance and, eventually, an M&A mandate. The swap is often the least profitable line item and the one that consumes the most capital, and it is priced as part of a whole. That is a legitimate commercial judgement, and it is also exactly the judgement that must be made explicitly, in a governance forum, with the XVA charge visible — because a relationship rationale applied without a number attached is how banks accumulate uncollateralised long-dated exposure without noticing.

Banks and dealers: small effective exposure, concentrated and correlated risk

Interbank counterparty risk is mostly collateralised away, and what remains is the risk that collateralisation itself fails. Trades between dealers run under an ISDA Master with a Credit Support Annex, typically with a zero threshold, a small minimum transfer amount, daily variation margin in cash, and — where the Uncleared Margin Rules bite — segregated initial margin computed under ISDA SIMM. A large share of the standardised flow is cleared through CCPs. The effect is that current exposure between dealers is small and the residual exposure is essentially the margin period of risk: what can the position move by between the last margin exchange and the completion of a close-out?

That is a modest number per counterparty and a dangerous number in aggregate. Dealers face each other in the same instruments, hedge with each other, and hold correlated positions, so a dealer default is never an isolated event. The counterparty that fails is also the counterparty everyone else was relying on for a hedge, and the collateral that must be liquidated is the collateral everyone else is trying to liquidate simultaneously. This is the systemic contagion channel that the entire post-2008 build — clearing mandates, margin rules, the CVA capital charge — was designed to interrupt. Lehman Brothers is the canonical demonstration: roughly 900,000 derivative transactions across 6,500 legal entities had to be closed out, and the litigation over close-out valuations ran for more than a decade.

Hedge funds and leveraged accounts: the risk is the speed of the move, not the rating

A leveraged fund is heavily collateralised, frequently unrated, and can still destroy a prime brokerage business in four days. The instinct trained on corporate lending fails badly here. The fund posts initial margin and daily variation margin; on the credit dimension the bank looks well protected. But initial margin is sized to cover the plausible move over the margin period of risk, and everything depends on whether that sizing is live, whether the portfolio is concentrated, and whether the margin model re-rates when the portfolio changes shape.

Archegos is the reference case, and it is the reference case for a specific reason. The failure was not that the family office was a poor credit — its credit assessment was largely beside the point. The failure was that the prime brokers applied static percentage margin on total return swaps that never re-rated as the portfolio concentrated into a small number of illiquid single names, and that the aggregate position across dealers was invisible to each of them. When the names gapped, the margin was mechanically insufficient, the positions could not be liquidated without moving the market further against the liquidator, and losses accumulated in days. Map it to Stages 3, 4 and 7: the limit framework did not constrain concentration, the pre-deal check did not price the incremental concentration, and the daily exposure cycle produced a number that was stale relative to the speed of the move.

The right mental model for leveraged accounts is liquidity-adjusted, concentration-aware and dynamic. Margin must be a function of the portfolio's liquidation profile, not a fixed percentage of notional. Limits must bite on concentration and on the ratio of position size to average daily volume, not only on notional. And the exposure number must be produced faster than the market can move, which is a technology statement as much as a risk one.

Pension funds, insurers and asset managers: creditworthy and, in 2022, unable to post

These counterparties are usually excellent credits and they generate two problems that have nothing to do with default. The first is directionality. A pension scheme hedging its liabilities, or an insurer hedging guaranteed annuities, holds long-dated receive-fixed interest-rate swaps and inflation swaps in very large notionals, all pointing the same way. There is no offsetting flow inside the netting set, so netting achieves little and the exposure profile is one-sided and enormous. The second problem is the one 2022 taught expensively.

The crucial distinction is between a counterparty that cannot pay and a counterparty that cannot post — and they are not the same failure. A counterparty that cannot pay is insolvent: its assets are worth less than its liabilities, and the loss is a credit loss. A counterparty that cannot post is solvent and illiquid: it holds ample assets, but not in a form it can deliver as eligible collateral, in the required currency, within the required settlement window. Its net worth may be rising even as it fails the margin call. The credit analysis says the name is safe; the collateral analysis says the name will default on a technicality that the ISDA documentation treats as a full Event of Default.

In September 2022 that distinction stopped being theoretical. The UK gilt sell-off forced liability-driven investment pension funds into forced selling of the very gilts whose price fall had generated the collateral calls, and the Bank of England intervened. In the same year, European utilities faced tens of billions of euros of margin calls as gas and power prices spiked, and several governments extended emergency liquidity lines to solvent companies that simply could not source the cash. In both cases the counterparties were creditworthy. What they lacked was collateral capacity. The operational consequence is that onboarding and credit review for these counterparties must assess the liquidity buffer, the eligible-collateral inventory, the currency mix and the settlement operating model — not only the balance sheet.

Sovereigns, supranationals and central banks: low probability, enormous exposure, textbook wrong-way risk

Sovereign counterparties are frequently uncollateralised as a matter of policy rather than negotiation, and the exposures are very large. A sovereign or a supranational development bank will often decline to post collateral on principle, and the bank accepts it because the probability of default is low and the relationship is strategic. Central banks are a further step again. The result is a small number of counterparty relationships carrying disproportionate uncollateralised notional and very long tenors.

The wrong-way risk here is the classic textbook case, and it is specific rather than general. A bank holding a cross-currency swap with an emerging-market sovereign, where the swap gains value for the bank as that sovereign's currency depreciates, has an exposure that grows precisely as the sovereign's capacity and willingness to pay deteriorate — because currency collapse and sovereign distress are the same event viewed from two angles. Specific wrong-way risk is the version where the exposure is legally or structurally tied to the counterparty's own credit; general wrong-way risk is the statistical version, where exposure and credit quality are merely correlated. Sovereign FX exposure is the case that makes both definitions concrete.

CCPs: a counterparty by construction, a concentration by design

A CCP is technically a counterparty, and the entire point of it is to be one. When a trade is cleared, the original bilateral contract is replaced through novation with two contracts facing the CCP, so the bank's counterparty is no longer the dealer on the other side but the clearing house. The CCP protects itself with a default waterfall: the defaulter's initial margin, the defaulter's default fund contribution, a tranche of the CCP's own capital ("skin in the game"), the surviving members' default fund contributions, assessment rights over surviving members, and finally recovery tools. Exposure to a CCP is treated favourably in capital — cleared trades with a qualifying CCP attract a low risk weight — because the margin, the segregation and the waterfall are judged to make default remote.

The uncomfortable question is what the regime has actually done to the shape of systemic risk. Clearing did not remove counterparty risk; it concentrated it into a small number of systemically critical institutions and converted a diffuse credit problem into a concentrated liquidity problem. A member's exposure to a CCP includes not only its own margin but a contingent obligation to fund other members' failures through the default fund and assessment rights. The infrastructure built to solve "too big to fail" is itself, in every meaningful sense, too big to fail. That is not an argument against clearing. It is an argument for measuring, limiting and stress-testing CCP exposure as deliberately as any bilateral one.

Counterparty type Typical collateral arrangement Exposure profile Dominant risk What the bank charges for it
Corporates (NFC−/NFC+) None or one-way; no CSA below threshold Long-dated, rising then amortising, one-directional Default and wrong-way risk on the hedged commodity Full CVA, FVA and KVA; the largest XVA charge in the book
Banks and dealers CSA, zero threshold, daily VM, SIMM IM Short effective exposure; MPoR gap only Concentration and correlated systemic contagion Small CVA; MVA on posted IM; capital on the residual
Hedge funds and leveraged accounts Heavy IM plus daily VM, often intraday Small on paper; gaps violently on concentration Speed of move and liquidation concentration, not rating Financing spread and margin; limits do the real work
Pension funds, insurers, asset managers CSA, often VM-only, large notionals Very long-dated, one-directional, poorly netted Collateral capacity — cannot post rather than cannot pay CVA plus a liquidity and concentration premium
Sovereigns and supranationals Frequently uncollateralised by policy Very large, very long, uncollateralised Wrong-way risk and single-name concentration CVA constrained by relationship pricing; capital-heavy
CCPs Full IM, VM and default fund contribution Small direct exposure; contingent mutualised exposure Concentration in critical infrastructure; liquidity calls Low risk weight; default fund capital charge; MVA

The three lines of defence, and why the second line must be able to say no

A control that can be overruled by the person it constrains is not a control. The three-lines model is the industry's answer to that observation, and in counterparty risk it is enforced more literally than almost anywhere else in the bank, because the first line has both the information advantage and the revenue incentive.

The first line owns the risk it creates. Coverage bankers, sales and the trading desks originate the exposure, and the model holds them accountable for it — not merely for producing revenue but for operating within the limits, capturing the trade correctly, and escalating what they see. First-line risk ownership is a real obligation: the desk that books the trade owns the consequence of having booked it.

The second line is independent, and independence has a precise operational meaning. Credit risk, market risk and model validation sit outside the business, report through a Chief Risk Officer who reports to the board rather than to the head of the division, and — the load-bearing detail — their compensation does not depend on deal flow. A credit officer whose bonus moves with the desk's revenue is a first-line function wearing a second-line title. The independence requirement is what makes it possible for a credit officer to decline a profitable trade from a strategic client on a Friday afternoon and still hold the job on Monday.

The third line is internal audit, which does not assess exposures but assesses whether the first two lines did their jobs. Audit tests whether limits were genuinely enforced, whether breaches were escalated within the stated timeframe, whether model validation findings were remediated, and whether the evidence supports the story. It reports to the audit committee of the board, outside the executive chain entirely.

Line Who What they own Reports to Compensation linked to revenue
First Coverage, sales, trading, XVA desk Origination, pricing, booking, hedging, operating within limits Business heads Yes
Second Credit risk, market risk, model validation Limit setting, approval, model approval, independent challenge Chief Risk Officer No
Second (finance) Product control, Finance, capital reporting The CVA reserve, fair-value validation, capital calculation Chief Financial Officer No
Third Internal audit Assurance that lines one and two worked as designed Board audit committee No
Figure 14 Three lines, three verbs — and only the second line can say no.

Read Figure 14 for the veto markers rather than for the boxes. Three functions can stop a trade outright: Credit Risk can refuse the exposure, Legal can refuse to confirm that netting is enforceable, and Model Validation can suspend the model that produces the number. Every other function in the diagram influences, prices or reports. That asymmetry is the whole design — a control function that can only advise is a reporting function with strong opinions, and the distinction becomes visible only under commercial pressure, which is exactly when it matters.

The third column is nearly empty, and that is accurate rather than a drafting shortcut. Internal Audit tests whether the first two lines did what they said they would. It does not run a parallel risk process, it does not approve trades, and it arrives after the fact. Banks that staff the third line as though it were a second second-line end up with two functions doing the same review and nobody testing whether the review works.

Delegated authority and the escalation path

Authority in credit is delegated in a grid, not granted in a role. The grid maps the size and shape of an exposure onto the seniority required to approve it: notional or PFE bands, tenor bands, counterparty rating bands, and product categories, crossed against approval levels running from the individual credit officer through a credit committee to the board risk committee. A ten-year uncollateralised swap with a BBB− corporate takes an exposure request well above what a single officer may sign. The Helvetia trade illustrates the mechanism cleanly: at 91% utilisation of the EUR 75m group PFE limit, the trade fits — but the utilisation level is itself a trigger for review, and any further request from the same group escalates automatically because there is no headroom left to consume.

Escalation is defined by trigger, timeframe and destination, not by judgement. A limit breach is logged automatically, notified to the desk and the credit officer the same day, and must be either cured — by unwinding, hedging, novating or taking collateral — or formally approved as a temporary excess by someone in the grid with authority for the new level. What is not permitted is silent persistence. A breach that sits open for weeks without a decision, or is quietly resolved by adjusting the limit to fit the exposure, is the specific governance failure supervisors look hardest for.

The XVA desk is a deliberate hybrid, and the discomfort is the design

The XVA desk is a trading desk in the first line that produces a number the second line depends on, and the industry has never made that comfortable. It prices the counterparty credit component of every new trade, takes the resulting CVA, FVA and KVA onto its own book, and hedges the sensitivities in the market — credit, FX, rates. That makes it a first-line risk taker with a P&L. But the number it produces, the CVA reserve, is simultaneously a risk measure, a capital input and an accounting entry.

The tension is managed by making the CVA reserve an accounting number that Finance owns and validates under IFRS 13. IFRS 13 requires fair value to reflect counterparty credit risk, which means the CVA is not an internal risk opinion the desk may adjust to suit its position — it is an audited component of the balance sheet, computed under a validated model, reviewed by product control and tested by external auditors. The desk trades it; Finance owns the level; model validation owns the method. No single function can move the number alone. That three-way lock is what makes it acceptable for a revenue-generating desk to produce a control-relevant figure.

The XVA desk changes the internal economics of the bank in a way worth stating explicitly. Before XVA desks existed, counterparty risk was a cost dumped on the originating desk or absorbed centrally with no owner. Concentrating it in one book creates a transfer-pricing mechanism: the originating desk pays the XVA desk the charge at inception, and thereafter the counterparty risk is the XVA desk's problem to hedge or hold.

That has two consequences. The originating desk's P&L becomes clean market-risk P&L, which is what it is competent to manage. And the counterparty risk becomes visible in one place, managed by people who do nothing else, with a mandate and a limit. The bank pays for this in complexity — an internal charge everyone argues about — and buys clarity about who owns what.

Two constraints sit outside the bank's own appetite, and they cannot be negotiated

Everything described so far — the limit hierarchy, the delegated authority grid, the escalation triggers — is the bank's own statement of what it is willing to lose. A bank can change any of it by decision. Two further constraints are not the bank's to set, and confusing them with internal appetite is a category error that surfaces at exactly the wrong moment.

The large exposures regime is a hard ceiling, not a risk appetite

Prudential rules cap what a bank may be exposed to in respect of a single client or group of connected clients, expressed as a percentage of its own funds. In the European framework this sits in the Capital Requirements Regulation; in the United States the equivalent is the single-counterparty credit limit applying to the largest institutions. The construction that matters is group of connected clients: entities linked by control, or so economically dependent that the difficulty of one would cause the difficulty of the other. That definition is a judgement the bank must make and defend, and it is exactly the judgement the legal entity hierarchy in Part 4 exists to support.

The regime is stricter still between large institutions, and derivative exposures count. Exposures to other global systemically important institutions face a tighter cap, on the reasoning that interconnection between the largest firms is the specific channel through which a single failure becomes a systemic one. Counterparty credit exposure computed under SA-CCR feeds this calculation, which means a decision taken for capital reasons — an internal model, a netting recognition, a compression cycle — moves a hard regulatory ceiling as a side effect.

The practical consequence is that a bank can be well inside its own appetite and out of room. A relationship the credit committee is entirely comfortable with can be closed to new business because the group breached a connected-clients ceiling somewhere else in the institution. Front-office teams experience this as arbitrary, and the correct response is not to explain the rule but to surface the remaining headroom in the pre-deal check alongside the internal limit, so that the constraint is visible before a price is quoted rather than after.

Credit risk mitigation extends beyond collateral, and each tool has a different owner

Collateral dominates the discussion in this module because it dominates the market, but it is one of several ways to reduce a counterparty exposure and it is not always available. Helvetia posts nothing and will not agree to a CSA; the question of what else can be done is a real one, and it is answered by a set of tools that sit with different functions.

Tool What it does Who owns it The catch
Parent or affiliate guarantee Substitutes a stronger entity's credit for the trading entity's Credit Risk and Legal Only as good as the guarantee's enforceability and the parent's own credit
Credit insurance A third-party insurer indemnifies the loss on default Credit Risk, with Legal on policy terms Recognition for capital is conditional and the insurer becomes a new concentration
Single-name or index CDS Transfers default and spread risk to the protection seller XVA desk No liquid market for most corporate names, and index hedging leaves basis risk
Break clauses Allow termination at set dates, shortening the effective tenor Credit Risk in negotiation The counterparty prices the option, and exercising it damages the relationship
Third-party novation Transfers the trade and its exposure to another dealer Trading and Legal Requires a willing transferee, which is scarcest precisely when it is needed
Netting and compression Reduces the exposure without touching the risk position Operations and Legal Depends entirely on a clean netting opinion holding

The distinction that matters for capital is between mitigation that is recognised and mitigation that merely helps. A guarantee from a rated parent, documented in a form the supervisor accepts, substitutes the guarantor's risk weight and reduces the capital requirement. An informal comfort letter from the same parent reduces nothing, however reassuring the credit committee finds it. The difference is entirely a documentation question, which is the same lesson Stage 2 taught about netting and the same one Part 5 draws from Lehman.

Information barriers: the credit officer knows something the salesperson is not allowed to hear

The classic conflict on a trading floor is between advisory information and trading activity. In counterparty risk the conflict is different and less discussed: the bank's own credit view of its client is material, non-public and commercially sensitive. The bank knows things about the counterparty that the counterparty's own market has not priced, and it knows them because the counterparty told the bank in confidence during onboarding and annual review. That knowledge cannot be allowed to leak into how the bank trades, and it also cannot be allowed to leak out.

Three specific situations recur, and each is genuinely uncomfortable. The first: a client is moved to the internal watchlist because credit has seen a deteriorating covenant position, and the following morning a salesperson quotes that client a ten-year swap without knowing. If the salesperson is told, the salesperson's behaviour towards the client will change in ways the client can read, and the bank has effectively disclosed its private credit assessment. If the salesperson is not told, the bank writes long-dated uncollateralised exposure to a name it has internally downgraded. Neither outcome is clean, and the resolution is structural: the watchlist status does not reach the salesperson as a fact but reaches the limit system as a constraint, so the pre-deal check simply returns a smaller number or a refusal, without explanation.

The second: the XVA desk is about to buy CDS protection on a client while the coverage banker is preparing a pitch to that same client. The desk is hedging a legitimate exposure. The banker is doing legitimate business. But the desk's hedging activity is a signal about the bank's credit view, and in a thin single-name CDS market it is a signal the market can read. The barrier here works in both directions — the desk must not learn from coverage that the client is about to announce a transaction, and coverage must not learn from the desk that the bank is buying protection.

The third is not a conflict of information but a conflict of position, and it deserves to be stated plainly: buying CDS protection on your own client is commercially and ethically fraught. The bank is simultaneously advising a company, lending to it, and holding a position that pays out if it fails. Several banks have been publicly criticised for exactly this, and some restrict or prohibit it by policy on named relationship clients, hedging instead with sector indices or proxy names that carry basis risk but no single-name signal. The trade-off is real and there is no costless answer: index hedging leaves residual idiosyncratic risk that the CVA capital charge will penalise, while single-name hedging manages the risk precisely and creates a relationship exposure that is difficult to defend if it becomes public. The decision belongs in a governance forum, documented, not on a trading desk at 4pm.

The mechanisms that enforce all of this are the restricted list, the watchlist and surveillance. The restricted list is visible: it names entities in which the bank has restricted its own or its employees' activity, and its existence is known across the firm. The watchlist is confidential: it names entities where the bank holds material non-public information, it is maintained by Compliance, and knowing that a name is on it is itself sensitive. The asymmetry creates the practical problem — the control must operate on people who are not permitted to know the control is operating, which is why the enforcement has to live in systems rather than in instructions.

The barriers are enforced by configuration, not by trust. Five mechanisms carry the weight.

Entitlements and role-based access control ensure that credit commentary, watchlist status and internal rating rationale are visible only to the second line; the front office sees a limit and a charge, not the reasoning behind them.

The restricted list is enforced at the point of order entry and at pre-deal check, so a blocked activity fails automatically rather than relying on someone remembering.

Watchlist access control is logged and reviewed, because the audit question is not only who acted on the information but who looked at it.

Communication surveillance covers email, chat and voice with lexicon-based and behavioural detection, calibrated to flag credit-view leakage as well as the more familiar market-abuse patterns.

Personal account dealing rules require pre-clearance against the restricted list, so employees cannot trade on what the firm knows.

Every one of these produces an audit trail, and the trail is the evidence the supervisor will ask for.

The teams: nine functions, ten stages, and a system behind each one

Counterparty risk touches more distinct teams than any other single discipline in a corporate and investment bank, which is why the handoffs are where it fails. The table below is the working map. Read it alongside the stage numbers from Part 2 — each team owns specific stages, and the boundary between two teams is always a stage boundary, because that is where a data object passes from one system to another.

Team Division Core role Stages owned Primary technology
Coverage / Sales Front office, client-facing Owns the relationship; originates the trade; delivers the price 1, 4, 5 CRM, pricing screen, pre-deal check API
Trading desk Front office, markets Prices and hedges the market risk; books the trade 5, 6 Trade capture, front-office risk system
XVA desk Front office, central Prices and hedges the counterparty risk; owns the XVA book 5, 9 XVA engine, Monte Carlo grid, CDS trading
Credit Risk Second line, independent Assesses the counterparty; sets and enforces limits; approves excesses 1, 3, 4, 10 Rating engine, limit system, watchlist
Collateral Operations Operations Calls, receives, values and disputes margin; manages eligibility 8 Collateral platform, SWIFT, tri-party links
Model Validation Second line, independent Independently challenges and approves exposure and XVA models 7, 9 Model inventory, backtesting framework
Finance / Capital Second line, control Owns the CVA reserve under IFRS 13; computes RWA and capital 5, 9 General ledger, capital engine, SA-CCR calculator
Legal and Documentation Control Negotiates ISDA, Schedule, CSA; obtains netting opinions 2, 10 Document management, netting-opinion library
Regulatory Reporting Operations / control Reports trades and lifecycle events to repositories 6, 7 Reporting engine, ISO 20022 XML, repository links

Coverage owns Stage 1 and Stage 4, and its system is the pre-deal check. The relationship manager brings the counterparty in, gathers the onboarding pack, and later stands in front of the client with a price. What Coverage does not own is whether the price is right or whether the exposure fits — those come back through the limit system and the XVA engine as constraints. The salesperson's screen is the point where three upstream systems collapse into a single number and a green or red light.

Credit Risk owns Stage 1 jointly, Stage 3 outright, and the veto at Stage 4. It converts the onboarding pack into an internal rating and a PD, converts the rating and the relationship into a limit grid, and then lives inside the limit system, which is the enforcement point for everything it has decided. At Stage 10 it reappears as the owner of the distressed-name process. Its technology is unglamorous and load-bearing: a rating engine, a limit system holding utilisation by legal entity and netting set, and a watchlist with restricted access.

Legal owns Stage 2, and Stage 2 determines whether the numbers everyone else computes are legally real. Whether exposure may be measured net rather than gross depends entirely on whether close-out netting is enforceable in the counterparty's jurisdiction, and that is a legal opinion sitting in a document library, mapped to a netting-set identifier in the risk system. The fifth of the module's six ideas — the loss happens at close-out but it is decided at documentation — is a statement about this team.

The XVA desk owns Stage 5 and Stage 9, and it is the only team that appears on both sides of the pricing conversation and the hedging conversation. At Stage 5 it computes the charge from the exposure profile, the credit curve and the funding cost. At Stage 9 it manages the resulting book, hedging credit spread, FX and rates sensitivity, and absorbing the P&L when spreads move — which is where the third illustrative figure of the module bites: a CVA reserve moving from EUR 4.2m to EUR 9.6m on a downgrade is a EUR 5.4m realised loss with no default in sight.

Collateral Operations owns Stage 8, and its clock is the shortest in the module. It runs the daily valuation, issues and answers margin calls, applies eligibility rules and haircuts, moves securities and cash through custodians and tri-party agents, and manages disputes. In stress it runs intraday. Its systems talk SWIFT and increasingly ISO 20022, and its failure mode is not analytical but operational — a call not made, a dispute not resolved, a settlement that missed a cut-off.

Model Validation and Finance own the credibility of the numbers rather than the numbers themselves. Validation independently challenges the exposure simulation and the XVA models, and backtests forecast exposure against realised mark-to-market — the evidence the supervisor requires before granting or maintaining IMM permission. Finance owns the CVA reserve as an audited IFRS 13 fair-value component and produces the capital numbers under SA-CCR or the IMM. Between them they turn a model output into something the board and the auditor can rely on.

  1. Why is the bank described as the only party that can price the netting set, and in what precise sense did Archegos disprove the broader version of that claim?

  2. Helvetia is uncollateralised, unrated by the CDS market and hedging for ten years. Why does the bank accept an exposure with those characteristics?

  3. How does counterparty type affect the CVA charge, and why can an uncollateralised BBB− corporate cost more than a leveraged unrated fund?

  4. Distinguish a counterparty that cannot pay from one that cannot post, and explain why the distinction mattered in 2022.

  5. Why must the second line's compensation be independent of deal flow, and what is the delegated authority grid for?

  6. The XVA desk sits in the first line but produces a number the second line relies on. How is that tension contained?

  7. Why can the watchlist not simply be shown to the sales desk, and how is the control implemented instead?

  8. What is the argument against buying single-name CDS protection on a relationship client, and what is the cost of the alternative?

3

  1. What did Archegos disprove?

  2. Why can an uncollateralised BBB− corporate cost more CVA than a leveraged unrated fund?

  3. Why does the bank accept Helvetia's uncollateralised exposure?

4The exposure number is only as good as the weakest system that touched it

Counterparty risk is the most computationally demanding number a bank produces and the one with the longest data supply chain. It starts in a legal document, passes through a trade capture system, a curve builder, a Monte Carlo engine, a netting-set mapping, a collateral platform and a capital calculator, and arrives on a supervisor's desk as a single figure. Any one of those steps can be wrong without anything looking wrong. This part follows the number through the machinery.

Ask what the capability must do before asking who sells it

Vendor names age badly and functions do not. Every bank running a derivatives business — a global dealer, a regional universal bank, a specialist financing house — needs the same set of functional blocks, whatever it calls them internally and whoever built them. Naming the functions first makes the architecture legible, makes buy-versus-build a decision rather than an accident, and makes it possible to look at any bank's system estate and see what is missing.

Seven blocks carry the load, and each has a single defining obligation. Client and credit onboarding must turn a legal entity into an approved, rated, identified counterparty with a limit. The legal document store must hold the ISDA Master, the Schedule, the CSA and the netting opinion, and — the part banks underestimate — must expose their economic terms as structured data rather than as PDFs. Limit management must hold every limit and its live utilisation, and must be able to answer a question about a trade that has not happened yet. The pricing and XVA engine must produce the charge. Exposure simulation must produce the distribution — EE, EPE, PFE — across every future date for every netting set. Collateral and margin must compute what is owed today, call it, receive it, value it and dispute it. Capital and regulatory reporting must turn the exposure into EAD, RWA and a report a supervisor accepts.

Cutting across all seven is a band of capabilities that belong to no single block and break all of them when they fail. Market and credit reference data supplies curves, volatilities, fixings, spreads and the proxy sector curves that stand in where no CDS trades. The legal entity hierarchy and LEI resolve who the counterparty actually is and what group it belongs to. The netting-set definition — a derived object, discussed at length below — tells every downstream system which trades may be offset against which. Model risk governance holds the model inventory, the validation status and the approved parameters. The immutable audit trail records what every system said and when. Grid and cloud compute supplies the raw arithmetic. Model validation and backtesting supply the evidence that the numbers deserve belief. Stress testing supplies the numbers the models do not produce on their own.

Block What it must do Core data it owns What breaks without it
Onboarding Identify, rate and approve a counterparty; assign LEI and internal ID Client static data, internal rating, PD, hierarchy No counterparty exists to trade with; no PD for CVA
Legal documents Hold ISDA, Schedule, CSA, netting opinions as structured terms Thresholds, MTA, eligible collateral, netting enforceability Exposure must be measured gross; capital rises sharply
Limit management Hold limits and live utilisation; answer pre-deal queries in seconds Limit grid, utilisation by netting set and legal entity No enforceable control; Stage 4 becomes advisory
Pricing and XVA Compute CVA, FVA, KVA and the total charge for a new trade XVA reserves, credit curves, funding curves, sensitivities Counterparty risk is taken but never priced
Exposure simulation Generate paths and produce EE, EPE, PFE per netting set Exposure profiles across every future date No limit measure, no CVA input, no IMM capital
Collateral and margin Value, call, receive, haircut, segregate and dispute margin Positions, eligibility, haircuts, call and dispute history Collateral terms exist on paper and nowhere else
Capital and reporting Produce EAD, RWA, and regulatory and trade-repository reports SA-CCR inputs, RWA, reported trade records The bank cannot demonstrate solvency or compliance
Figure 15 The functional architecture: seven blocks and the capabilities that cut across all of them.

The buy-versus-build line is a map of where the bank believes it has an edge

In practice the seven blocks are filled by a mix of vendor platforms, industry utilities and internal builds, and the split is not random. Onboarding and client lifecycle management is dominated by vendor platforms, often shared with the KYC and AML function, because the requirements are regulatory and near-identical across banks. Document and agreement management likewise: the ISDA Master is a standard form, industry initiatives have pushed towards digitised and structured representations of agreement terms, and a bank gains nothing from a proprietary way of storing a document everyone else stores the same way.

Limit systems, exposure engines and XVA libraries split differently, and the split is instructive. Vendor engines for exposure simulation and XVA exist and are used, including by very large institutions. So do substantial in-house builds — a number of major dealers run proprietary XVA and exposure platforms developed over a decade or more, because the model choices, the calibration and the netting treatment are considered part of how the bank competes. The same institution will frequently buy the collateral platform and build the pricing library, and that combination is not inconsistent.

Collateral management is largely a vendor and utility space, and the tri-party agents sit at its centre. Collateral platforms handle valuation, call generation, eligibility testing, haircut application and dispute workflow. Tri-party agents — the large international central securities depositories and major custodian banks — hold and allocate collateral between parties so that neither has to trust the other's segregation. Regulatory reporting is even more clearly a utility space: trades are reported to trade repositories, and the reporting layer between the bank and the repository is a commodity that no bank derives advantage from writing itself. Market data is bought, from the major curve, price and credit-spread vendors, with internal work concentrated on curve construction and proxy-curve methodology rather than on the raw inputs.

The organising principle is simple and worth stating as a rule. Banks buy where the function is a commodity defined by a regulator or an industry standard — reporting formats, collateral messaging, document storage, static data — because the specification is written by someone else and conformance, not creativity, is the objective. Banks build where the function encodes a judgement they consider proprietary — the exposure model, the XVA pricing library, the limit policy engine, the proxy credit curve methodology — because those are the places where two banks looking at the same trade legitimately produce different numbers. The buy-versus-build line is therefore not a procurement outcome. It is a map of where the institution believes its edge lies, and it can be read that way from the outside.

Block Typical approach Why
Onboarding and client lifecycle Buy Requirements are regulatory and common; shared with KYC and AML
Legal document store Buy, with internal structuring layer The documents are standard; the extraction of terms is not
Limit management Build or heavily configure Limit policy is the bank's own risk appetite made executable
Pricing and XVA engine Build at large dealers; buy elsewhere Model and calibration choices are where valuation opinions differ
Exposure simulation Mixed; vendor engines and major in-house platforms Regulator approves the model, not the vendor; IMM raises the stakes
Collateral and margin Buy, plus tri-party agents High-volume operational processing on standard message formats
Capital and regulatory reporting Buy for reporting; build for capital calculation Report formats are prescribed; capital aggregation is bank-specific

Follow one number front to back and the failure modes become obvious

Figure 16 Data lineage: one wrong netting-set identifier becomes a wrong capital number overnight.

A single trade generates a chain of data objects, each computed from the last, and the chain runs overnight without a human in it. Take the Helvetia cross-currency swap. It is booked on the afternoon of 15 March 2026 and by 06:00 the next morning it has contributed to a limit utilisation, a CVA reserve, a collateral position, a capital number and a regulatory report. Nobody looked at it in between. Following that chain is the fastest way to understand where counterparty-risk data actually breaks.

It starts in trade capture, where the economics and the identifiers are created together. The trader books USD 500m against EUR 460m, ten years, fixed-for-fixed, and the system generates a unique transaction identifier, attaches the counterparty's legal entity and LEI, assigns a product classification, and — critically — resolves the trade to a netting set by looking up which ISDA Master governs this legal entity pair. That resolution is the highest-consequence lookup in the chain and it is invisible on the screen.

From there the trade enters the overnight cycle described in Part 2 Stage 7. Market data close produces the curves. The trade population snapshot picks up every live trade including this one. Monte Carlo path generation produces thousands of scenarios for EUR/USD and the two rate curves. Pricing along paths values the swap at every future date on every path. Aggregation collapses those values into netting sets — this is where the netting-set identifier does its work — and produces EE, EPE and PFE profiles. Collateral terms are applied, which for Helvetia means nothing is applied because there is no CSA. The results are distributed.

And then four systems consume the same profile for four different purposes. The limit system compares peak PFE of EUR 68m against the EUR 75m group limit and records 91% utilisation. The XVA engine integrates EPE against the proxy hazard-rate curve and produces the CVA reserve. The collateral system records that no margin is callable. The capital engine computes SA-CCR: adjusted notional EUR 460m, FX supervisory factor 4%, add-on EUR 18.4m, replacement cost zero at inception, EAD of 1.4 × (0 + 18.4) = EUR 25.8m. The reporting engine assembles the ISO 20022 XML record for the trade repository. All figures illustrative only.

Stage System Data object created What consumes it downstream
6 — Booking Trade capture Trade record with UTI, LEI, netting-set ID, product class Risk engine, reporting engine, collateral system
7 — Curve build Market data and curve engine Discount, forward and credit curves at close Monte Carlo path generation, XVA pricing
7 — Simulation Exposure engine Path-level valuations across future dates Netting-set aggregation
7 — Aggregation Exposure engine EE, EPE, PFE per netting set Limit system, XVA engine, capital engine
3, 4 — Limits Limit system Utilisation record and breach flag Pre-deal check, credit escalation, dashboards
5, 9 — XVA XVA engine CVA, FVA, KVA reserves and sensitivities Finance ledger, hedging book, FRTB-CVA capital
8 — Collateral Collateral platform Margin call amount, eligibility and haircut decision SWIFT instructions, custodian, tri-party agent
9 — Capital Capital engine EAD, RWA, capital requirement Regulatory returns, capital planning, limit calibration
6, 7 — Reporting Reporting engine ISO 20022 XML trade and lifecycle report Trade repository, supervisor, reconciliation

The cascade principle is that an error upstream reaches every downstream system, and in counterparty risk it does so silently. In a settlement failure, the failure announces itself: a payment does not arrive, someone calls. In exposure computation, nothing announces itself, because every output of a wrong input is a perfectly plausible number. A netting set assigned to the wrong legal entity produces an exposure profile, a CVA reserve and an EAD. They are all internally consistent, all pass every sanity check, and all wrong. The error surfaces at close-out, when the netting the bank assumed turns out not to be enforceable against the entity that actually defaulted, or at a supervisory review two years later. That is the counterparty-risk-specific version of the cascade, and it is why lineage documentation and reconciliation are controls rather than paperwork.

Lineage is a plumbing diagram, and its value is the same as a plumbing diagram's. Nobody draws one because the water is flowing. It is drawn so that when something is wet in the basement, the number of places worth looking is small and known. A bank without lineage debugs an exposure discrepancy by asking six teams what they think happened.

BCBS 239 is not a reporting rule, it is an architecture rule

In 2013 the Basel Committee published its principles for effective risk data aggregation and risk reporting, and the reason it did so is the most damning sentence in post-crisis supervision. During the crisis, banks could not answer basic questions about their own exposures fast enough for the answers to be useful. Boards asked what the firm's total exposure to a named counterparty was, across every product and every legal entity, and the honest answer took days. Some of the institutions that failed did not know what they were owed by the institutions that failed alongside them. BCBS 239 exists because supervisors concluded that a risk number produced too slowly to act on is not a risk number.

Four of the principles do the real work in this domain. Accuracy and integrity requires that aggregation be largely automated, because manual intervention is where errors and delays both live. Completeness requires that the aggregation covers all material risk across all business lines and legal entities — the word that bites is all, since partial coverage is exactly what a patchwork estate produces. Timeliness requires that the numbers arrive quickly enough to matter, with the standard set by stress conditions rather than by a quiet Tuesday. Adaptability requires that the bank can answer a question nobody anticipated, which is a demand for a flexible data model rather than for another report.

The supervisory test that matters is a single question asked under pressure. What is our aggregate exposure to this named group, across all products and all entities, right now, and how confident are you in it. A bank that can answer that within hours has, by construction, solved the netting-set derivation problem, the legal-entity hierarchy problem, the reconciliation problem and the lineage problem described in the preceding sections. A bank that cannot has one or more of them unsolved, whatever its architecture diagram claims.

Data quality is four dimensions, and each fails differently

Data quality in counterparty risk is not a general aspiration; it decomposes into four measurable dimensions with different controls and different consequences. Treating it as one undifferentiated concern is why data quality programmes so often produce dashboards instead of improvements.

Dimension What it means here A concrete control What it costs when it fails
Completeness Every trade, netting set and collateral agreement is present in the risk population Trade-count and notional reconciliation between front office and risk, with a zero-tolerance threshold Exposure is understated by exactly the trades that were missed, and nothing in the output looks wrong
Accuracy Field values are right, especially netting set, CSA link, legal entity and maturity Validation rules at capture, plus counterparty portfolio reconciliation Wrong netting, wrong limit, wrong margin call, wrong capital — the full cascade
Timeliness The number is available before the decision that depends on it Batch completion monitoring against the 06:00 deadline, with alerting on the critical path Controls operate on stale data; limits and margin calls are enforced against yesterday
Consistency The same trade means the same thing in every system that holds it Cross-system field-level comparison, not just totals Reconciliation breaks that are individually small and collectively unexplainable

The governance question is harder than the technical one, and it is usually the one left unanswered. When the exposure number is wrong because a reference table was stale, the failure belongs to nobody in particular: the risk function consumed the data, the front office entered it, operations maintained the table, and technology ran the pipeline. Banks that make progress here do one specific thing — they name a single accountable owner for each critical data element, and that owner sits in the business rather than in technology. Ownership by committee is how a data quality issue survives three remediation programmes.

The trade record and the netting set: the field-level anatomy

Everything downstream is computed from a trade record, and a surprising amount of it is computed from fields nobody on the desk looks at. The economic fields — notional, rate, maturity — are checked constantly because they drive the P&L. The structural fields — netting-set identifier, CSA identifier, legal entity, product classification — are checked rarely because they drive nothing the trader sees. They drive capital.

Field What it holds What it drives downstream
Trade identifier (UTI) Globally unique ID for this transaction, shared with the counterparty Repository reporting, reconciliation, lifecycle events, dispute resolution
Counterparty legal entity The specific legal person on the other side, not the group Netting-set assignment, limit consumption, close-out rights
LEI The 20-character global identifier of that legal entity Regulatory reporting, group hierarchy roll-up, concentration measurement
Netting-set identifier Which pool of trades this may be offset against Exposure aggregation, EAD, CVA, collateral, capital — everything
CSA identifier Which collateral agreement, with which thresholds and MTA Collateral calls, MPoR assumption, exposure reduction, MVA
Product classification ISDA taxonomy code for the instrument type Clearing eligibility, SA-CCR asset class and supervisory factor, UPI
Notional and currency, leg 1 USD 500m in the running example Valuation, SA-CCR adjusted notional, reporting, exposure scale
Notional and currency, leg 2 EUR 460m in the running example Same, plus the FX exposure that dominates the profile
Trade date 15 March 2026 Reporting deadline, lifecycle sequencing, inception CVA booking
Maturity date 15 March 2036 Exposure profile length, SA-CCR maturity factor, tenor limits
Payment schedule Coupon dates, frequency, day count, fixings Cash flow projection along every simulated path
Collateral eligibility flag Whether and what collateral applies to this trade Whether exposure is measured collateralised or uncollateralised
Clearing status Cleared, uncleared, or clearing-eligible but not cleared Capital treatment, CCP exposure, EMIR obligation
Internal desk and book Which desk owns it and which book it sits in XVA transfer pricing, P&L attribution, limit ownership

A netting set identifier is a postcode, not a description. It carries no information about what the trade is or how risky it looks, and it determines everything about where the trade is delivered — which pile it lands in, what it is allowed to offset against, and which legal agreement governs it when things go wrong. Two trades that look identical on a term sheet and carry different netting set identifiers are, to every system downstream, in different countries.

The netting set is derived, not captured — and that is where the capital goes

Every other field on that list is entered or generated at booking. The netting-set identifier is inferred, and inference is where mistakes live. A netting set is the pool of trades that may legally be offset against one another on default, and its boundary is defined by a legal agreement: an ISDA Master Agreement between two specific legal entities, in a jurisdiction where close-out netting has been confirmed enforceable by a netting opinion. It is not a system concept that someone typed in. It is a legal fact that a system must correctly reflect.

Getting it right is worth a great deal of capital, and getting it wrong in the conservative direction is expensive but invisible. If a bank cannot demonstrate that netting is enforceable — because the opinion is missing, or stale, or does not cover the specific product or entity type — it must measure exposure gross. Recall the illustrative whole-portfolio cascade used elsewhere in this module: gross positive mark-to-market of EUR 142m falls to EUR 38m after close-out netting, to EUR 12m after variation margin, and to approximately zero after initial margin. The step from EUR 142m to EUR 38m is netting alone, and it is entirely a function of documentation being correct and correctly represented in the risk system.

Banks lose more capital to conservative netting-set mapping than to any other data problem, and they lose it quietly. The pattern is familiar: an agreement exists but the opinion has not been refreshed, so the trades are split into single-trade netting sets. A subsidiary trades under the parent's Master through a guarantee, but the system maps it to its own entity. A cleared trade and an uncleared trade with the same counterparty sit in separate netting sets, correctly, but a subsequent trade is mapped to the wrong one. Each error produces a higher exposure and a higher capital number, nobody complains because the direction is prudent, and the bank pays for it every quarter.

Legal entity resolution: LEI helped, and did not solve it

The same client group appears under many names, and the counterparty is a legal entity while the credit view is about a group. Helvetia Chemicals AG may trade through a Swiss parent, a Dutch financing subsidiary and a US operating company, each with its own LEI, its own balance sheet and — critically — its own ISDA Master. Netting operates at the legal entity pair. Limits and concentration measurement operate at the group. Both views must be simultaneously correct, which requires a maintained hierarchy linking every entity to its ultimate parent.

LEI made the identification problem tractable and left the hierarchy problem largely open. A twenty-character LEI unambiguously names a legal entity and removes the ambiguity of matching on name strings, which used to mean deciding whether "Helvetia Chemicals AG", "HELVETIA CHEM AG" and "Helvetia Chemicals Aktiengesellschaft" were the same counterparty. That is a real gain. But LEI coverage is driven by reporting obligations, so entities outside scope may not have one; parent relationship data exists but is not complete or uniformly maintained; and the hierarchy a bank needs for credit purposes — economic dependency, guarantee structures, sponsor relationships — is a credit judgement that no registry encodes. Banks therefore run their own hierarchy on top of LEI, and that hierarchy is a maintained data asset with an owner, not a static reference table.

Reconciliation exists because three systems hold the same trade differently

The front office, risk and collateral systems each hold a version of the trade, built for a different purpose, and they will not naturally agree. Front office holds it for pricing and P&L, at trade granularity, updated in real time. Risk holds a snapshot for simulation, sometimes with the trade decomposed or approximated for computational tractability. Collateral holds it at portfolio level, aggregated by CSA, with its own valuation. Add the counterparty's own record and the trade repository's record and there are five versions.

Deduplication and reconciliation are therefore continuous processes, not periodic clean-ups. Internal reconciliation compares trade counts, notionals and valuations between front office, risk and collateral each day and investigates the breaks. External portfolio reconciliation — a risk-mitigation requirement for non-cleared derivatives under EMIR — compares the bank's record with the counterparty's, and it is the mechanism that catches a missing trade or a differing maturity before it becomes a dispute. Repository-level reconciliation matches the two sides' reports on the shared UTI, which is precisely why the UTI must be agreed between the parties rather than generated independently.

Identifiers: five strings that determine whether anything reconciles

Counterparty risk depends on identifiers more than most disciplines because an OTC derivative has no natural name. An equity has an ISIN. A ten-year cross-currency swap between two named entities has nothing except what the reporting framework constructs for it, which is why identifier standards are load-bearing infrastructure here rather than administrative detail.

Identifier What it names Where it bites in counterparty risk
LEI A legal entity, globally, in 20 characters Netting-set boundaries, group concentration, every regulatory report
UTI One transaction, agreed between both parties Two-sided repository reconciliation; a mismatched UTI is an unreconciled trade
UPI The product type, so a bespoke swap is classifiable Aggregation by product, clearing eligibility, supervisory factor mapping
ISIN A security Collateral eligibility, haircuts, tri-party allocation, settlement
CSA / netting-set ID The bank's own pointer to a legal agreement Whether exposure is measured net or gross; the highest-leverage internal field

The distinction that trips people up is between an identifier the market assigns and an identifier two parties must agree. LEI, ISIN and UPI come from registries — the bank looks them up. The UTI is generated by one party under an agreed rule and adopted by the other, which means a reporting framework that does not specify clearly who generates it produces two reports of the same trade that never match. The netting-set identifier is purely internal and appears in no external message at all, which is exactly why it receives less scrutiny than its consequences justify.

Three integration patterns, three different things you can control

Batch is the legacy default, and its weakness in this discipline is specific rather than general. The overnight cycle runs from roughly 18:00 to 06:00, produces a full set of exposure profiles, and distributes them at dawn. It is efficient, it is auditable, and it means the exposure number in front of a credit officer at 15:00 was computed from last night's positions and last night's prices. On a quiet day, that is fine. In a fast market it means the bank is enforcing limits against a picture that is up to twenty-four hours stale — which was exactly the Archegos problem. The margin was based on yesterday, the positions moved today, and the arithmetic never caught up.

API integration answers a different question: what would happen if I did this trade? The pre-deal check at Stage 4 is a synchronous call from the pricing screen into the limit and exposure systems, asking for the incremental exposure of a hypothetical trade against the existing netting set, and it must return in seconds because a salesperson is on the phone. This is architecturally harder than it looks, because computing an incremental PFE properly means re-aggregating the netting set with the new trade included, not adding a standalone number to a stored one. Most banks solve it with pre-computed path-level exposures for the existing portfolio, against which the candidate trade can be valued quickly.

Event streaming changes the question again: the exposure updates as the world does. Trades, prices, collateral movements and lifecycle events arrive as events on a stream, and the exposure and limit state are updated continuously rather than rebuilt nightly. It is a substantially harder engineering problem — full revaluation cannot be done per event, so the design depends on incremental updates, cached path valuations and approximation with periodic full recomputation as a control.

Pattern Latency Typical use in counterparty risk Principal weakness
Batch Overnight, 18:00 to 06:00 Full exposure profiles, capital, regulatory reporting Number can be up to 24 hours stale when it matters most
API Seconds Pre-deal check, incremental exposure, price and charge Requires pre-computed state to be fast enough to be useful
Event streaming Sub-second to minutes Intraday exposure, limit monitoring, margin triggers Approximation risk; harder to reconcile and to audit

Moving from nightly batch to streaming exposure is not a performance upgrade. It changes what is controllable.

A limit is only enforceable at the frequency at which the exposure is measured. A bank with an overnight cycle can have an intraday limit policy written in a document, but it cannot enforce one, because on any given afternoon it does not know the number. The control exists on paper and not in the system.

So the business case for real-time exposure is not "the same number, faster". It is a new class of control: intraday limits that actually bite, margin triggers that fire on the move rather than the following morning, and concentration alerts that arrive while the position can still be reduced. The engineering cost is real and the honest justification is capability, not speed.

Buy or build is the wrong question when the answer is agree

This part has so far framed the platform decision as buy versus build, and for several of the most expensive functions in counterparty risk that framing is simply wrong. Reconciliation, confirmation, margin messaging, dispute management and compression share a property that changes the economics completely: their value comes from the counterparty using the same one. A reconciliation service that only your bank uses reconciles nothing. This is the third option, and it is neither buying a product nor building one — it is agreeing on a shared representation and then processing against it.

The Common Domain Model attacks reconciliation at the root

Everything this part has said about reconciliation assumes that two firms will represent the same trade differently and then discover the differences afterwards. The ISDA Common Domain Model refuses that premise. It is a standardised, machine-readable representation of trades, lifecycle events and the processes that act on them, expressed so that two firms processing the same event produce the same result by construction rather than by comparison. A partial termination, a novation, a compression tear-up or a margin call has one definition, one set of fields and one expected outcome.

The distinction between automating reconciliation and removing the need for it is the whole point. A bank can invest heavily in matching engines, tolerance rules, exception queues and break-ageing dashboards, and every euro of that spend is a euro spent managing a problem rather than removing it. A shared model removes the divergence that creates the breaks. It is a slower, more political and considerably more valuable path, and it is the only one that makes the reconciliation layer smaller rather than more efficient.

Digital documentation turns capital-relevant clauses into queryable fields

The trade-record anatomy section established that the CSA threshold, the minimum transfer amount, the eligible collateral schedule and the downgrade triggers all drive exposure and capital. In most banks those parameters live in a signed PDF and reach the systems that need them by being read by a person and typed into a screen. That is the actual state of the art for a set of fields that determine a regulatory capital number.

Digital documentation platforms negotiate and store these agreements as structured data. ISDA Create and comparable services allow the Schedule and the Credit Support Annex to be negotiated field by field, with the executed result available as a machine-readable object rather than as a scan. The immediate operational benefit is the elimination of a manual keying step. The larger benefit is a question that becomes answerable for the first time: which of our agreements contain a ratings downgrade trigger at BBB− or above, and what would we have to post if we were downgraded tomorrow. AIG's failure turned on precisely that question, and most banks still cannot answer it from a database.

The utility layer is a distinct architectural pattern, not a procurement choice

A group of functions in this chain are mutualised across the market rather than bought or built. Portfolio reconciliation services compare two firms' views of the same portfolio and surface the differences. Margin call messaging and dispute management platforms carry calls, responses and disputes in a common format so that both sides see the same workflow. Trade matching and confirmation platforms establish an agreed record at the point of execution. Compression services run the multilateral cycles described in Part 2.

None of these could be built usefully by one bank alone, and that is the defining feature of the category. The buy-versus-build analysis asks whether a capability is a source of competitive advantage. The mutualisation question is different and simpler: does this function only work if my counterparty is doing the same thing. Where the answer is yes, the correct strategy is to join the network that already exists and to spend the saved effort on the decision engine, which is where the bank's own judgement actually lives.

Approach What it standardises Why it beats reconciliation What it demands of the bank
Common Domain Model The representation of trades and lifecycle events Divergence never occurs, so there is nothing to reconcile Rebuilding internal models around an external standard
Digital documentation Legal agreement terms as structured fields Capital-relevant parameters stop being manually keyed Renegotiating or re-papering existing agreements
Market utilities The workflow between two firms Both sides see one record instead of two Accepting an external process and its release cycle
Distributed ledger The state itself, held once rather than twice There is only one copy to be wrong Legal, operational and governance change at market scale

Distributed ledger deserves an honest paragraph rather than an enthusiastic one

The theoretical fit is genuinely good and the practical record is genuinely thin. A shared ledger removes reconciliation by removing the second copy, and a self-executing collateral agreement removes the dispute by removing the discretion. Those are real answers to real problems described in this part. After a decade of pilots, consortia and proofs of concept, very little of the bilateral derivatives lifecycle actually runs this way, and the honest reasons are legal enforceability, governance of the shared infrastructure, and the difficulty of migrating a live market rather than any technical limitation.

The part of the idea that landed is the standardisation, not the ledger. The Common Domain Model emerged from this body of work and is being adopted because it delivers most of the benefit without requiring the market to agree on who runs the infrastructure. That is worth noticing as a general pattern: when a technology narrative underdelivers, the durable residue is usually the data standard it forced everyone to agree on.

The reason standardisation is hard here is political rather than technical, and it is worth being precise about the mechanism. The reconciliation layer is expensive, but its cost is spread across operations budgets in every bank simultaneously and appears on no single decision-maker's page. Removing it requires competitors to agree on a shared representation, to rebuild internal systems against it, and to accept an external release cycle governing something on their critical path. Each firm's rational move is to wait for the others to go first, which is why standards in this market advance in the wake of regulation rather than ahead of it — the rule supplies the deadline that no participant would have imposed on itself.

Compute is the binding constraint, and that is unusual for a risk system

Most risk calculations in a bank are cheap; counterparty exposure is not, and the difference is structural. Market risk value-at-risk revalues today's portfolio under a few hundred or a few thousand scenarios. Counterparty exposure revalues the portfolio at every future date on every simulated path, for every netting set, and then does it again for every sensitivity needed for CVA hedging and FRTB-CVA capital.

The cost model is a product of four terms and it grows uncomfortably fast. Paths multiplied by time steps multiplied by trades multiplied by counterparties. Ten thousand paths, a hundred future time points, a hundred thousand trades and a few thousand counterparties gives an order of magnitude of individual valuations in the hundreds of billions per nightly run — and each of those valuations is not a lookup but a pricing call, which for anything with optionality is itself expensive. Full brute-force revaluation of a real derivatives portfolio in a twelve-hour window is not achievable, and every technique described below exists to avoid it.

American Monte Carlo, also called least-squares Monte Carlo, solves the early-exercise problem without nesting simulations. Products with optionality — Bermudan swaptions, callable structures, anything with an exercise decision — need a value conditional on the state of the world at a future date, and the naive approach is a simulation inside a simulation, which is computationally impossible at portfolio scale. The technique instead regresses the future payoff on observable state variables across the existing paths, producing an approximate conditional expectation as a function rather than as a nested simulation. One set of paths, a regression at each exercise date, and the continuation value falls out. It is the standard method for exposure on early-exercise products.

Adjoint algorithmic differentiation is the single largest efficiency gain in this field, and it deserves to be understood rather than named. Start with the problem it solves. To hedge a CVA reserve, or to compute the sensitivities that SA-CVA demands, the bank needs the derivative of the value with respect to every risk factor — every point on every curve, every volatility, every credit spread. That can easily be a thousand or more factors. The obvious method is bumping: shift one factor slightly, revalue the whole portfolio, take the difference, repeat. One thousand risk factors means one thousand full revaluations, and if a single revaluation of the exposure profile takes hours, the calculation cannot be finished.

AAD computes all of those sensitivities in roughly the cost of a small multiple of a single valuation, regardless of how many risk factors there are. The intuition is that a valuation is a sequence of elementary arithmetic operations, and the chain rule can be applied to that sequence. Running through it forwards gives the value. Running back through the same recorded sequence in reverse, propagating derivatives, gives the sensitivity of the output to every input in one pass — because each intermediate result's contribution to the final answer is accumulated on the way back rather than rediscovered by re-running the whole computation. The cost of the reverse pass is a small constant multiple of the forward pass, typically in the region of two to five times, and that constant does not grow with the number of inputs. A calculation that scaled as one thousand valuations now scales as a handful. This is what makes daily XVA sensitivity computation and sensitivity-based CVA capital practical at all, and it is a genuine change in what is possible rather than an incremental optimisation.

GPU acceleration attacks the same problem from the hardware side. Monte Carlo path generation and path-wise valuation are close to ideally parallel — each path is independent of every other path — which is exactly the shape of computation a graphics processor is built for. Moving path generation and pricing kernels onto GPUs delivers large speed-ups over general-purpose processors for the numerically intensive inner loop, at the cost of rewriting pricing libraries to fit the execution model.

Cloud bursting handles the shape of the demand curve rather than its height. The compute requirement is not flat: month-end capital runs, quarter-end reporting, regulatory stress tests and ad-hoc "what if this counterparty defaults" analyses spike far above the nightly baseline. Sizing an on-premises grid for the peak means paying for idle capacity most of the year. Bursting the peak into cloud capacity is a natural fit, subject to the real constraints of data residency, model governance and the requirement that a cloud-computed regulatory number be as auditable as an on-premises one.

Technique What it attacks Rough order of gain Principal cost
American Monte Carlo Early-exercise products needing conditional values Turns an infeasible nested simulation into a feasible single pass Regression approximation error; requires validation
AAD Sensitivity computation across many risk factors From one valuation per factor to a few valuations in total Pricing library must be rewritten or instrumented
GPU acceleration The path generation and valuation inner loop Large, on the parallel portion of the workload Code rewrite; specialised skills; hardware estate
Cloud bursting Peak demand at month-end and in stress Removes the need to own peak capacity Data residency, governance, auditability of results

Why AAD is the reverse pass and not the forward pass is worth one more paragraph, because it explains the asymmetry.

Algorithmic differentiation can be run in two directions. Forward mode propagates the derivative with respect to one chosen input through the whole computation, so its cost is proportional to the number of inputs — the same scaling problem as bumping, just with better accuracy. Reverse (adjoint) mode propagates the derivative of one chosen output backwards to all inputs, so its cost is proportional to the number of outputs.

Counterparty risk has thousands of inputs — every curve point, every volatility, every spread — and a handful of outputs: the value, the CVA, the exposure profile. Many inputs, few outputs is exactly the configuration where reverse mode wins overwhelmingly, which is why the field standardised on adjoint rather than forward differentiation.

The intuition behind adjoint differentiation is worth having, because it explains a difference of two orders of magnitude. Bumping asks, for each input in turn, what happens to the answer if I nudge this one — so a portfolio with two thousand risk factors needs two thousand extra valuations. Adjoint mode runs the calculation once and then walks it backwards, asking what each input contributed to the answer that was produced. It is the difference between weighing every ingredient separately to find out what made the cake heavy, and reading back through the recipe once. The cost of all sensitivities becomes a small multiple of the cost of one valuation, rather than a multiple of the number of inputs.

Immutable audit and model risk governance: the record is the control

Exposure and limit records must be append-only, because the alternative is a system in which a breach can be made never to have happened. If a limit utilisation of 103% can be corrected in place to 98% after the position is reduced, then the record no longer shows that the breach occurred, that it went unescalated for a day, or that someone decided not to report it. The correction is not fraud in most cases — it is usually a well-intentioned data fix — and it is still a governance failure, because it destroys the evidence on which the control depends.

The rule is that corrections are new records, not edits. An amended exposure figure is stored as an amendment, with a timestamp, an author, a reason and a link to what it superseded, and the original remains readable. A limit breach that is subsequently cured shows as a breach followed by a cure, not as an absence. Supervisors test exactly this: they ask for the breach log for a period, compare it with the escalation records and the credit committee minutes, and look for breaches that appear in one and not the others. A clean-looking log with no breaches in a volatile quarter attracts more attention than a log with breaches that were properly escalated.

Model risk governance is the second half of the same idea, applied to the methods rather than the records. An exposure model or an XVA model is approved by a validation function independent of the developers and the users, which reviews the theoretical soundness, the implementation, the calibration and the limitations, and imposes conditions or capital add-ons where it finds weaknesses. Approval is not permanent. Models are periodically re-validated, and material changes require re-approval.

Backtesting is the specific evidence supervisors demand for exposure models, and it tests a forecast rather than a valuation. The model predicts a distribution of future exposure — EE, EPE, PFE at various horizons. Backtesting compares those forecasts against the exposure that was actually realised as time passed and the mark-to-market evolved. If realised exposure exceeds the 97.5% PFE forecast far more often than 2.5% of the time, the model understates the tail and the finding must be remediated or capitalised. Under the IMM this is not optional: supervisory approval to use internal models for counterparty credit risk capital is conditional on a functioning backtesting programme, and failing it can cost the permission.

The model inventory is itself a system, and treating it as a spreadsheet is a recognised failure. It must record every model in use, its version, its owner, its validation status and expiry, its known limitations, its approved scope of use, and the systems and reports that depend on it. That last item is the one that matters operationally: when validation finds a problem with a curve-construction model, someone must be able to answer immediately which exposure numbers, which XVA reserves and which regulatory submissions were computed with it. Without a dependency map, that question takes weeks.

An immutable log is a ship's logbook rather than a spreadsheet. Entries are appended in ink, corrections are made as new entries that reference the old ones, and nothing is ever quietly amended after the fact. That property is worthless on a good day and decisive on the day a supervisor, a litigant or an internal investigation asks what the bank knew and when it knew it — a question no editable record can answer credibly.

Four messaging layers, four clocks, and the gaps between them

Figure 17 Four messaging layers, four clocks. Counterparty risk lives in the gaps between them.

Different parts of the trade lifecycle speak different protocols, and each protocol was designed for a different problem in a different decade. This is not untidiness that will eventually be resolved. It reflects genuinely different requirements — execution needs speed and terseness, legal representation needs structural completeness, settlement needs guaranteed delivery, and regulation needs prescribed fields.

FIX carries execution flow, and it is built for latency. The Financial Information eXchange protocol is the standard for order routing and execution reporting in electronic markets: quotes, orders, fills, cancellations. Its messages are compact and its timescale is milliseconds. It describes what was traded and at what price, and it is deliberately not the place where the full legal and economic structure of an OTC derivative lives.

FpML is the layer that matters most for counterparty risk, because a swap has no ISIN and must be described structurally. Financial products Markup Language is an XML standard for representing OTC derivatives: the legs, the notionals, the schedules, the fixing conventions, the day counts, the business day adjustments, the optionality, the parties. A ten-year fixed-for-fixed EUR/USD cross-currency swap cannot be identified by a code, because it is not an instrument that exists independently of the two parties who agreed it — it is a bespoke contract. FpML is how that contract moves between systems and between firms in machine-readable form: confirmations, portfolio reconciliation files, valuation exchanges, and the payload behind much post-trade processing. When people say a derivative "has no security identifier", FpML is the answer to the question that raises.

SWIFT carries the money and the securities, on a settlement clock. Collateral and cash movements use SWIFT message types: the MT54x family for securities settlement instructions and confirmations, the MT2xx family for cash transfers, and MT569 for the tri-party collateral and valuation statement that tells a party what collateral is held, how it is valued and what haircuts were applied. These messages are reliable, auditable and slow by front-office standards — their timescale is the settlement cycle, and their guarantee is delivery rather than latency. The migration to ISO 20022 is replacing the MT syntax with richer structured XML across payments and securities, which matters here because it carries more data per message and reduces the truncation and free-text problems that have historically caused collateral breaks.

The regulatory layer is the newest and the most prescriptive. EMIR REFIT moved trade reporting to ISO 20022 XML with a specified field set, and reports go to authorised trade repositories rather than to a counterparty. The format is not negotiable, validation is performed by the repository, and rejected reports are a compliance issue rather than an operational inconvenience.

Layer Standard Carries Timescale
Execution FIX Orders, quotes, fills, cancellations Milliseconds
Product representation FpML (XML) The full structure of an OTC derivative contract Trade date to T+1; reconciliation daily
Settlement and collateral SWIFT MT54x, MT2xx, MT569; ISO 20022 Securities movements, cash transfers, tri-party statements Hours to settlement cycle
Regulatory ISO 20022 XML to trade repositories Trade, position and lifecycle reports under EMIR and equivalents Within the reporting deadline, typically T+1

Operational resilience: the daily cycle has no second chance

If the overnight batch fails, the bank does not get a chance to catch up before the market opens. It starts the trading day with yesterday's exposure profiles. The limit system holds stale utilisations, so the pre-deal check at Stage 4 either blocks legitimate business or waves through trades against limits that no longer reflect reality. The collateral system cannot compute today's margin call, so the bank does not issue the call it is entitled to and cannot properly evaluate the one it receives. The capital number is yesterday's. And on the specific morning where all of this matters most — the morning after a large market move, when counterparties are under strain and margin calls are large — the bank does not know what it is owed.

That asymmetry is what "mission critical" means in this context, and it is why the vocabulary needs to be precise. Redundancy means duplicated components, so no single hardware, network or process failure stops the system — but redundancy alone only guarantees that a spare exists. Automatic failover means the switch to that spare happens without human intervention and within a defined time, which is the part that turns redundancy into availability. Disaster recovery is the capacity to restore service at a geographically separate site after the loss of the primary one, which addresses the failure modes redundancy within a single site cannot. Recovery time objective is the maximum acceptable duration of the outage, and recovery point objective is the maximum acceptable amount of data loss expressed as time — how far back the restored state may be. Business continuity is broader than all of them: the plan by which the business function continues, including manual workarounds, delegated authorities and communication, when the systems are unavailable.

Concept What it guarantees Why it matters on a volatile morning
Redundancy A spare component exists for every single point of failure A failed grid node does not stop the exposure run
Automatic failover The switch to the spare happens without human action Nobody has to be awake at 03:00 to save the 06:00 delivery
Disaster recovery Service can be restored at a separate site A site-level loss does not mean a day without exposure numbers
Recovery time objective (RTO) The maximum tolerable outage duration Sets whether the cycle can still deliver before market open
Recovery point objective (RPO) The maximum tolerable data loss, measured in time Determines whether trades booked late yesterday survive
Business continuity The function continues by other means Margin calls still get issued, manually, at reduced scope

The distinction that matters most is between redundancy and tested failover. A bank with a fully redundant estate and an untested failover procedure has bought hardware, not resilience. Failover has to be exercised, under realistic conditions, with the actual overnight cycle running, because the failure modes that matter are the ones that only appear when the switch happens under load — a dependency that was never replicated, a credential that only existed on the primary, a downstream system that does not accept a connection from the secondary address. Regulators have made operational resilience an explicit supervisory theme precisely because the paper architecture and the tested architecture diverge so reliably.

Five regulatory regimes, read as system specifications

Regulations in this field are usually summarised legally and are more useful read as architecture requirements. Each of the five below tells the bank to build something specific, hold specific data, and produce specific evidence. Reading them that way makes the system estate in the first half of this part look less like an accumulation of history and more like a set of answers.

EMIR and Dodd-Frank Title VII specify a reporting and reconciliation machine. Every derivative must be reported to a trade repository within the reporting deadline, typically the next business day, in a prescribed format — ISO 20022 XML under EMIR REFIT. That requires the bank to generate and agree a UTI with the counterparty, classify the product for a UPI, hold LEIs for both parties, capture the prescribed field set at booking rather than reconstruct it later, report lifecycle events as they occur, reconcile portfolios with counterparties at a prescribed frequency, and determine for each trade whether it falls under the clearing obligation. That last determination is a rules engine reading product classification, counterparty category and threshold status — for Helvetia, an NFC− below the clearing threshold, the answer is exempt, and the system must be able to demonstrate why.

CRR with SA-CCR and the IMM specifies a calculation with an audit trail. SA-CCR is a prescribed formula — EAD = 1.4 × (RC + PFE add-on) — and it requires a specific data set: adjusted notionals, supervisory factors by asset class, maturity factors, netting-set and hedging-set structure, and collateral terms. Every input must be reproducible and every intermediate step reviewable, because a supervisor may ask the bank to reproduce a single netting set's EAD from a specific date. The IMM permits the bank's own simulation instead, with EAD = α × effective EPE and α normally 1.4, but the price of that permission is a validated model, a backtesting programme and a governance apparatus that must be maintained continuously.

FRTB-CVA specifies a granularity that legacy engines frequently cannot reach. The revised CVA risk framework offers BA-CVA in reduced and full versions and SA-CVA, the sensitivity-based standardised approach that requires supervisory approval. SA-CVA needs CVA sensitivities to prescribed risk factors, computed at a specified granularity, daily. An engine built to produce a single CVA reserve number overnight cannot simply be asked for a thousand sensitivities as well. This is precisely where AAD stops being an optimisation and becomes the enabling technology, and it is the most common reason banks rebuild rather than extend their XVA platform.

UMR and ISDA SIMM specify a daily calculation, a segregation arrangement and a dispute process. In-scope counterparties must exchange variation margin, and initial margin was phased in from 2016 to September 2022, the final phase catching counterparties with average aggregate notional amount above EUR 8bn. IM must be segregated and cannot be rehypothecated, which is a custody and account-structure requirement as much as a calculation one. ISDA SIMM is the standard model, so the bank must run it daily on in-scope portfolios, exchange sensitivities in the prescribed format, compare its result with the counterparty's, and operate a defined dispute resolution process when the two disagree.

IFRS 13 turns the risk number into an audited accounting number, which changes who controls it. Fair value must reflect counterparty credit risk, so the CVA is a component of the balance sheet, subject to external audit, requiring documented valuation methodology, fair-value hierarchy classification and disclosure. The practical consequence for the system estate is that the XVA engine is not only a risk system. It feeds the general ledger, which means its outputs need reconciliation to accounting records, controlled change management and evidence that an auditor will accept — a materially higher bar than an internal risk model faces.

Regime What it demands of the system Where it usually hurts
EMIR / Dodd-Frank Title VII Repository reporting within T+1, UTI agreement, reconciliation, clearing determination Field completeness at booking; UTI mismatches between the two sides
CRR — SA-CCR and IMM Prescribed formula, full input data set, reproducible and auditable Netting-set and hedging-set structure; reproducing a past calculation
FRTB-CVA Daily CVA sensitivities at prescribed granularity Legacy engines cannot produce them; drives AAD adoption or rebuild
UMR / ISDA SIMM Daily IM calculation, sensitivity exchange, segregation, dispute process Segregated custody structures; reconciling SIMM inputs with counterparties
IFRS 13 CVA as an audited fair-value component with documented methodology Ledger reconciliation and audit evidence from a model-driven number

  1. Distinguish FIX, FpML and SWIFT, and explain why FpML is the one that matters most for counterparty risk.

  2. Why is the netting-set identifier the highest-leverage field in the trade record?

  3. Why is redundancy alone not enough for the overnight exposure cycle?

  4. Explain AAD in one paragraph and say why reverse mode rather than forward mode.

  5. What is specifically wrong with an overnight batch as the basis for limit control, and what does Archegos illustrate?

  6. Why do banks buy collateral platforms and build XVA engines?

  7. Why must exposure and limit records be append-only?

  8. Why can a legacy XVA engine typically not support SA-CVA, and what does the bank do about it?

  9. What did LEI solve, and what did it leave unsolved?

  10. What does BCBS 239 actually require, and why is it an architecture rule rather than a reporting rule?

  11. What problem does the Common Domain Model solve that a better reconciliation engine cannot?

  12. Why is adjoint differentiation such a large win for XVA?

4

  1. Why is the netting-set identifier the highest-leverage field?

  2. Which message standard matters most for counterparty risk?

  3. Why is redundancy alone not enough overnight?

5Every counterparty-risk failure is a failure to measure, to price, to collateralise or to document

The four preceding parts built the machine: what the risk is, how the lifecycle runs from onboarding to close-out, who does what across the ecosystem, and which systems carry the numbers. This part asks the only question that matters afterwards — how does it break, and how would you know before it broke? It sets out the six risks the discipline actually manages, the handful of indicators that separate a well-run counterparty book from a lucky one, three real failures that map exactly onto three stages of the lifecycle, the technologies that are changing the economics, and the specific place generative AI earns its keep. It closes with the six ideas worth carrying out of the module.

The six risks are one exposure seen from six angles, and each has a different owner

A bank does not have six counterparty risks; it has one exposure that can hurt it in six distinct ways. The Helvetia swap from Part 2 is a single contract with a single mark-to-market, and yet it can cost the bank money through a default, through a spread widening with no default at all, through a correlation between the two, through an inability to move collateral, through an error in the model that values it, and through a clause in a document signed years before the trade. Those are not overlapping descriptions of the same event. They are separate loss mechanisms with separate defences, separate owners and separate warning signs, and a bank that manages one well and another badly is not partially protected — it is exposed through whichever one it neglected.

The taxonomy matters because it assigns accountability. Each of the six sits with a different function, is measured by a different number, and is defended by a different control. Confusing them produces the most common organisational failure in this business: a risk that everyone agrees is serious and nobody actually owns. What follows takes each in turn, defines it precisely, and names the defence.

Default risk is the original risk and the smallest part of the modern problem

Default risk is the possibility that the counterparty fails and the bank loses the replacement cost of the trades it was in the money on. This is the risk the whole discipline was built to handle, and it remains the reference case. Its size is the product of three things the reader already has from Part 2: the exposure at the moment of default, the probability that default occurs, and the fraction not recovered afterwards. For Helvetia, that is an expected positive exposure of EUR 22.4m, a probability of default implied by a proxy spread of 180 to 220 basis points, and a loss given default of 60 per cent — the illustrative figures that produced the EUR 4.2m CVA charge.

What makes default risk deceptively simple is that it almost never happens, and when it does it is very large. A bank with a thousand corporate counterparties may go five years without a single derivative default and then lose more in one week than it earned in charges over the whole period. The distribution is not merely skewed; it is lumpy in a way that defeats intuition and, as the section on indicators argues below, defeats naive performance measurement entirely. The defences are the ones that operate before the event: an internal rating that is refreshed rather than filed, a limit framework that constrains the size of any single failure, and collateral that shrinks the amount at risk to something the bank can absorb.

Default risk is also the only one of the six that is genuinely transferred rather than merely reduced. A single-name credit default swap on the counterparty pays out on the credit event, and to that extent the bank has sold the risk to someone else. Everything else in this taxonomy is mitigated, structured around or operated against. That distinction — transferable versus manageable — is worth holding on to, because it explains why the CDS market matters far beyond its size and why the absence of a CDS on a name like Helvetia forces the bank into proxy hedging with an index, which reduces the risk without eliminating the basis.

CVA and spread risk is where the money is actually lost, and it is lost without any default

Between 2007 and 2009, roughly two-thirds of counterparty-credit losses across the banking system came from CVA mark-to-market moves rather than from actual defaults. That single fact reorganised the discipline and produced the Basel III CVA capital charge in 2010. It is worth restating plainly: most of what banks lost to counterparty risk in the crisis, they lost on counterparties that never failed. Credit spreads widened, the accounting value of the credit reserve rose, and the difference went straight through the profit and loss account under IFRS 13, which requires fair value to include counterparty credit risk.

Part 2's third illustrative figure set is the whole of this risk in one picture. At year three the swap has moved to plus EUR 41m in the bank's favour, Helvetia has been downgraded to BB+, its proxy spread has widened from 220 to 480 basis points, and the CVA reserve on this one trade has risen from EUR 4.2m to EUR 9.6m. The EUR 5.4m difference is a realised loss today, recognised in this quarter's earnings, and Helvetia has paid every cash flow on time. Nothing has gone wrong in any sense a credit officer would recognise. The bank has simply revalued a derivative it is short — because that is what a CVA reserve is.

Spread risk therefore behaves like a market risk and has to be managed like one. It moves daily, it has sensitivities to credit spreads, to interest rates, to FX and to volatility, and it can be hedged in the market. That is why the XVA desk described in Part 1 is a trading desk rather than a reporting function, and why the hedge-effectiveness ratio appears among the indicators below. It is also why the CVA capital framework has its own architecture — BA-CVA and SA-CVA under FRTB-CVA — rather than sitting inside credit capital. Two things follow for the reader. First, a bank can be perfectly right about who will default and still lose heavily on CVA. Second, a CVA book that is not hedged is not a conservative choice; it is an unhedged short credit position of institutional size.

Wrong-way risk is the correlation that makes every other number optimistic

Wrong-way risk is the tendency of the exposure to grow at exactly the moment the counterparty becomes less able to pay. The standard calculation treats exposure and default as independent: it takes the expected exposure profile from the simulation, multiplies by default probabilities from the credit curve, and adds up. If the two are correlated, that arithmetic understates the loss, because the cases where default happens are precisely the cases where the exposure was large. The convention splits it in two. General wrong-way risk arises from a macroeconomic link — a bank buying protection on a sovereign from a domestic bank of that sovereign, or an emerging-market corporate short its own currency. Specific wrong-way risk arises from a direct legal or structural link, the cleanest example being a counterparty that has pledged its own shares, or the shares of its parent, as collateral.

Right-way risk exists and is the mirror image, and mentioning it clarifies the concept. An airline hedging fuel is a better credit when fuel is cheap, which is exactly when the hedge is out of the money to the airline and the bank has no exposure. The exposure and the credit quality move in the helpful direction. Banks rarely give explicit credit for right-way risk in capital, but they do recognise it in appetite discussions, and the asymmetry — punished for wrong-way, not rewarded for right-way — is deliberate conservatism.

Wrong-way risk is defended structurally rather than numerically, because the numbers are not reliable enough to defend against. Correlation between a credit curve and a market factor is estimated from short and unstable histories, and the estimate matters most in the tail where the data is thinnest. So the practical controls are qualitative and hard: prohibit self-referencing collateral outright, apply concentration add-ons where the exposure and the counterparty share a driver, refuse or heavily haircut the trade where the link is direct, and run stress scenarios that force the market move and the downgrade to happen together rather than separately. Archegos, discussed below, was in part a wrong-way story — the collateral supporting the positions was the same set of concentrated equity holdings whose fall triggered the calls.

Collateral and liquidity risk is the risk created by the cure

Collateral does not delete risk; it converts credit risk into liquidity risk, and the liquidity risk lands on both sides. Once a CSA is in place, the bank's exposure to the counterparty's failure to pay is largely replaced by an exposure to the counterparty's failure to post, and by the bank's own obligation to post when the trade moves the other way. Part 2 Stage 8 traced the illustrative cascade for a whole-portfolio view of one counterparty: gross positive mark-to-market of EUR 142m, EUR 38m after close-out netting, EUR 12m after variation margin, approximately zero after initial margin. The residual credit risk is small. What has been created in its place is a daily cash obligation that scales with market volatility.

The residual exposure on a collateralised portfolio is the margin period of risk, and that is not a technicality. MPoR is the assumed gap between the last effective collateral exchange and the completion of close-out — ten business days for uncleared portfolios, five for cleared. Everything that happens to the mark-to-market inside that window is uncollateralised by construction. This is why initial margin exists at all, and why a collateralised book still generates a CVA charge, smaller but not zero.

The systemic form of this risk is procyclicality: everybody's collateral requirement rises at the same moment, for the same reason. Volatility spikes, initial margin models re-rate upward, variation margin calls follow the market move, and every participant needs cash on the same morning. That is the 2022 margin crisis in one sentence, and it is why the defence is not a credit control at all but a liquidity one — a stress test that asks what a bank and its counterparties would owe under a severe move, and whether the eligible assets on hand could be turned into the right currency of cash in time.

Model risk is the risk that the number defending you is wrong

Every figure on a counterparty-risk report is a model output, and none of them is observable in a market. No exchange quotes an expected positive exposure. Nobody prints a 97.5th percentile potential future exposure. These come out of a Monte Carlo simulation whose diffusion assumptions, volatilities, correlations, path count, time grid and collateral logic are the bank's own choices. If the calibration is stale, the exposure profile is wrong. If the collateral logic along the paths assumes margin arrives that would not arrive in stress, the profile is optimistic. If the proxy curve for a name like Helvetia — built from the European chemicals BBB sector curve because there is no public CDS — mis-ranks the credit, the CVA charge is mispriced from day one.

Model risk is unusual in that it has a fully institutionalised defence with a regulatory mandate behind it. Independent model validation sits outside the team that builds the model, challenges the assumptions, tests the implementation, and signs the model into an inventory with an approval, an owner, a review date and a list of known limitations. Backtesting compares the exposure the model predicted against what actually happened, at multiple horizons and confidence levels, and the exception rate is a supervisory item. Where a bank uses an internal model for capital under IMM, the entire apparatus is subject to supervisory approval and can be withdrawn.

The most dangerous model risk in this discipline is not a wrong formula but an unchallenged assumption. Formulas get tested. What survives is the thing nobody restated: that a collateral call is met the day it is made, that a proxy curve is representative, that a correlation observed in calm markets holds in a crash, that positions are liquidatable at the marks used in the simulation. Archegos was, among other things, a portfolio whose margin model had stopped describing the portfolio it was applied to.

Legal and operational risk decides how much of everything else you actually get to keep

Every mitigation described in this module is a contractual claim, and a contractual claim is worth what a court says it is worth. Close-out netting reduces the EUR 142m gross figure to EUR 38m only if the netting is enforceable in the counterparty's jurisdiction against its insolvency regime — which is what a netting opinion establishes, and which is why Part 2's Stage 2 takes four to sixteen weeks. Collateral protects the bank only if the security interest was validly created, if the assets were segregated where segregation was required, and if rehypothecation rights were what the bank thought they were. The Helvetia case is clean on this point precisely so that the reader can see what clean looks like: an ISDA Master with close-out netting and an unqualified Swiss netting opinion.

The operational half is the machinery that has to work under conditions nobody rehearsed. A close-out requires the bank to identify every trade with the defaulted entity — the correct legal entity, not the group — value them all on a defensible basis, serve notices within contractual deadlines, and document the process well enough to survive litigation years later. Lehman is the case study, and the numbers are the argument: roughly 900,000 transactions across 6,500 legal entities, with valuation disputes running for more than a decade.

This risk is the one most often mistaken for administration, and it is the one where the loss is decided. The trade generates the exposure, the model measures it, the collateral shrinks it — and the document determines how much of the shrinkage survives contact with an insolvency practitioner. A reference-data error in which legal entity a trade was booked against does not look like a risk issue on any normal day. On the day of a default it is the difference between netting and not netting.

Figure 18 Six risks, one exposure: the taxonomy is a map of who can control what.
Risk type What it is Who owns it Main defence
Default Loss of replacement cost when the counterparty fails Credit risk, second line Rating, limits, collateral, single-name CDS
CVA / spread Loss from credit spread moves with no default XVA desk Hedging with CDS and indices; CVA capital
Wrong-way Exposure grows as the counterparty weakens Credit risk with the trading desk Prohibitions, add-ons, joint stress scenarios
Collateral and liquidity Inability to post or to receive collateral in time Treasury and collateral operations Liquidity buffers, eligibility rules, stress testing
Model The measure itself is wrong Model validation, second line Independent validation, backtesting, model inventory
Legal and operational The mitigation fails when tested Legal and operations Netting opinions, documentation hygiene, close-out drills

How to tell a well-managed counterparty book from a lucky one

Ask a bank how its counterparty-risk management is performing and the tempting answer is "we have not lost money", which is not an answer at all. Counterparty losses are rare and lumpy by construction. A book can be run carelessly for years and show nothing, because no counterparty happened to fail; a book can be run impeccably and show a large loss, because one did. Outcome metrics in this discipline have almost no information content over the horizons on which people are judged. What can be measured is whether the machine that produces protection is working — whether the numbers are accurate, the limits bind, the collateral moves, the hedges offset and the volatility is contained. Those are process metrics, and in this discipline they are the primary evidence.

Five indicators do most of the work, and each has a characteristic way of being gamed. They are listed below with what good looks like, but the gaming column is the more useful one, because every metric in a bank is eventually managed rather than merely measured, and knowing the failure mode of a metric tells you which second question to ask.

Exposure backtesting exception rate

Backtesting compares what the exposure model predicted against what actually occurred, and the exception rate is the cleanest available evidence that the numbers mean something. For a given confidence level and horizon, the realised portfolio value should breach the predicted band about as often as the confidence level implies and no more. Persistent over-prediction is not "conservative"; it wastes limit capacity and overcharges clients. Persistent under-prediction means the limit framework is defending a line that is not where it is drawn. Supervisors look at this closely for banks on IMM, and the honest version reports by counterparty type, by asset class and by horizon, because an aggregate pass can conceal a systematic failure in one book.

Limit breach frequency and time-to-cure

A limit framework that is never breached is not necessarily well calibrated; it may simply be set too high to bind. The informative pair is how often limits are breached and how quickly breaches are cured. Frequent small breaches with same-day cures suggest limits that are doing their job at the margin. Rare breaches that persist for weeks suggest a framework that is treated as advisory. The pattern to watch for is repeated temporary increases on the same counterparty — a limit being raised to accommodate the business rather than the business being sized to the limit, which is one of the most reliable early indicators that appetite has quietly drifted.

Collateral dispute rate and ageing

A margin dispute is two institutions disagreeing about what a portfolio is worth, and the ageing of disputes is a direct read on operational quality. Disputes are normal; a portfolio of thousands of trades priced by two different models will not agree to the last euro, which is why CSAs contain dispute-resolution mechanics and minimum transfer amounts. What is not normal is a dispute that is still open after a month. An aged dispute means an uncollateralised exposure that nobody has decided to own, and a book with a long tail of them is carrying more risk than its collateral reports show.

CVA hedge effectiveness

The ratio of hedge profit and loss to CVA profit and loss over a period measures whether the XVA desk is actually offsetting the risk it holds. A ratio near one means the hedges track the reserve. A ratio that swings wildly means the desk is running basis — usually because it is hedging illiquid single names with an index, which is precisely the situation for a counterparty like Helvetia with no public CDS. Some basis is unavoidable and it is a legitimate business decision to run it. What is not legitimate is running it without measuring it, and the metric is gamed by choosing the measurement window that flatters the result.

Margin call settlement timeliness

The percentage of margin calls settled on the same day is a leading indicator of counterparty stress and of the bank's own operational capacity. A counterparty that starts settling late has usually developed a funding problem before it has developed a credit problem, which makes call timeliness one of the few genuinely forward-looking numbers in the discipline. It is also a control on the bank itself: a bank that cannot make its own calls on time in a volatile week has discovered the collateral and liquidity risk described above.

CVA volatility relative to earnings

A sixth measure sits above the others: how large is the quarter-to-quarter swing in the CVA reserve compared with the earnings of the businesses that generate it? This is the question a chief financial officer asks and the one that determines whether the XVA function is seen as a price or as a source of surprises. A CVA book whose volatility is a meaningful fraction of divisional earnings is either badly hedged or too large for the appetite, and either way it is a strategic issue rather than a technical one.

KPI What it measures What good looks like How it is gamed
Backtesting exceptions Whether the exposure model is accurate Exceptions in line with confidence level, stable across books Aggregating across books to hide one failing model
Limit breaches and cure time Whether limits bind and are enforced Some breaches, cured same or next day Setting limits high, or curing via temporary increases
Dispute rate and ageing Operational quality of the collateral process Disputes resolved within days, no aged tail Reporting count rather than value or age
CVA hedge effectiveness Whether the XVA book is actually hedged Hedge P&L tracking CVA P&L closely Choosing a flattering measurement window
Margin call timeliness Counterparty stress and own operational capacity High same-day settlement, stable over time Excluding disputed or partial calls from the count
CVA volatility vs earnings Whether the reserve is sized to appetite Small and stable relative to divisional earnings Netting reserve moves against unrelated releases

The scorecard is a smoke detector, not a fire alarm. It does not tell you the building is burning; it tells you the conditions in which a fire would spread are worsening. The temptation with a table like the one above is to wait for the bottom row to change before acting, and the bottom row in this business changes once a decade, discontinuously, and far too late to respond to.

Three real cases, three failure modes, three stages of the lifecycle

The three episodes below were not chosen because they are famous. They were chosen because between them they exhaust the ways this business fails. One is a failure to measure, one is a failure to document, and one is a failure to collateralise — and each maps onto specific stages of the ten-stage lifecycle from Part 2. Read together they are less a set of war stories than a structural argument about where the discipline is load-bearing.

Archegos, March 2021 — the exposure was never measured, so the limit never bound

Archegos Capital Management was a family office whose collapse cost Credit Suisse roughly USD 5.5bn, Nomura roughly USD 2.9bn and Morgan Stanley roughly USD 900m in a matter of days. It held very large, very concentrated positions in a small number of listed equities, and it held them not as shares but as total return swaps with a group of prime brokers. A total return swap gives the client the full economic exposure to a stock — the price movement and the dividends — in exchange for a financing payment, without the client ever owning the share. The bank owns the share as a hedge; the client owns the risk. That structure is entirely ordinary and it is the core of the synthetic prime brokerage business.

The first failure was that the margin never re-rated as the portfolio changed shape. The prime brokers applied static percentage margin to the swaps — a fixed haircut set at the outset of the relationship, of the kind that is perfectly adequate for a diversified book. As the positions grew and concentrated into a handful of names, the risk of the portfolio rose enormously while the margin requirement rose only in proportion to notional. A margin framework that does not respond to concentration is, in effect, a framework that assumes diversification it can no longer see. This is a Stage 7 failure — the daily exposure cycle was producing a number that did not describe the portfolio it was measuring — and a Stage 3 failure, because the limit framework was calibrated against that number.

The second failure was that no single prime broker could see the whole position. Each bank saw its own book and, on its own book, the exposure looked large but manageable. The aggregate — the same concentrated names, held through the same structure, financed by several banks simultaneously — was invisible to all of them. Because the exposure was synthetic, the disclosure obligations that attach to holding shares directly did not bite in the same way, so the position did not appear in the public filings that would ordinarily have alerted a credit officer to the scale of what a client was carrying elsewhere. The client knew its own leverage. No counterparty did.

The third failure was speed, and it is the one that turns a large exposure into a realised loss. When the underlying stocks fell, margin calls went out and were not met. Every prime broker held a hedge in the same small set of names, and every one of them needed to sell it into the same market at the same time. The first to liquidate got the best prices; the ones that hesitated, or that attempted a coordinated unwind, took the worst. The whole event ran its course in a few days. This is a Stage 4 point as much as a Stage 7 one: the pre-deal check that runs in seconds before each incremental trade never asked the question that mattered, which was not "does this trade fit the limit?" but "what does this portfolio look like if these names fall thirty per cent together?"

What changed afterwards is specific and durable. Synthetic prime brokerage books moved decisively toward dynamic margining, where the requirement is recalculated from the actual risk of the portfolio rather than fixed as a percentage of notional at onboarding. Concentration add-ons became standard, so that a portfolio's margin rises non-linearly as it concentrates into fewer names. And prime brokers began asking clients much harder questions about aggregate leverage across all their financing relationships, backed by a willingness to decline the business when the answers were unsatisfactory. None of these are new techniques. They are techniques that existed and were not applied to a client relationship that had been profitable for years — which is itself the lesson, and which connects directly to the fallacy set out in the previous section.

Lehman Brothers, September 2008 — the netting worked exactly as well as the paperwork

When Lehman Brothers filed for bankruptcy in September 2008, roughly 900,000 derivative transactions across some 6,500 legal entities had to be closed out. That is the number to hold on to, because it converts an abstract point about documentation into an operational reality. Every one of those transactions sat under some contract, with some counterparty, in some jurisdiction, and each had to be identified, valued, netted where netting was enforceable, and settled or litigated. The exercise did not take days. The litigation over close-out valuations ran for more than a decade.

The central finding is uncomfortably simple: close-out netting worked where the documentation was clean and failed where it was not. Where an ISDA Master Agreement was properly executed, where the Schedule said what the parties believed it said, where the netting opinion for that jurisdiction and that entity type was current and unqualified, and where trades had been booked against the correct legal entity, the mechanism performed as designed — the whole portfolio collapsed to a single net claim, and the gross-to-net reduction that Part 2 illustrated with the EUR 142m to EUR 38m cascade was realised in practice. Where any of those conditions failed, the counterparty found itself with gross claims and gross obligations, paying what it owed into an insolvent estate while queuing as an unsecured creditor for what it was owed.

The failures were rarely dramatic. They were unexecuted amendments, trades booked to a group entity rather than the trading entity, collateral held in accounts whose segregation status was not what the CSA implied, guarantees whose scope did not cover the products actually traded, and jurisdictions where the netting opinion carried qualifications that nobody had read since the file was opened. None of these looked like risk issues on an ordinary day. They looked like administration. On the day of the default they determined how much of the mitigation the bank actually got to keep — which is why this module treats Stage 2, the legal architecture, as a risk stage rather than a legal-department formality.

The valuation dispute was the second act, and it lasted far longer than the first. Closing out a portfolio requires determining a close-out amount, and the ISDA framework gives the determining party a defined but not mechanical procedure — obtaining quotations, or determining a commercially reasonable replacement value where quotations are unavailable. In September 2008 the markets in which those quotations would have been obtained were themselves dislocated. Two parties acting in good faith could arrive at very different numbers for the same portfolio, and the gap between them was worth enough to litigate for years. The lesson is that a close-out is not a calculation the bank performs; it is a position the bank will have to defend, with evidence, long after the event.

What changed is the reason this module's vocabulary looks the way it does. The industry adopted a series of protocols allowing counterparties to amend thousands of agreements multilaterally rather than one at a time, which is how the market later implemented benchmark transitions and resolution-stay requirements at scale. The G20 clearing mandate pushed standardised OTC derivatives to central counterparties, converting a web of bilateral close-outs into a managed default waterfall run by a single institution. Trade reporting under EMIR and Dodd-Frank Title VII gave supervisors a view of who was exposed to whom. And documentation hygiene became a discipline with an owner, a system of record and a data-quality metric, rather than a filing cabinet — the least glamorous of the changes and arguably the most valuable.

The 2022 margin crisis — the counterparties could pay, and still could not post

In 2022, European utilities faced tens of billions of euros in margin calls as gas and power prices spiked, and several governments extended emergency liquidity lines to keep them trading. The mechanics are worth stating carefully because the episode is routinely misdescribed as a credit event. A utility that has sold power forward — hedging generation it will actually deliver — is short the forward price. When prices rise violently, that hedge moves deep into loss on a mark-to-market basis, and every euro of that loss is called as variation margin, in cash, immediately. The economics of the underlying business have not deteriorated; the physical power the utility will deliver is worth more than ever. But the hedge and the physical position settle on completely different timetables, and the collateral call arrives years before the delivery revenue does.

This is the distinction the episode burned into the discipline: the ability to *pay* and the ability to *post* are different questions. A counterparty can be solvent, profitable, investment grade and entirely good for its obligations at maturity, and still be unable to find several billion euros of cash on a Tuesday morning. Every credit assessment described in Stage 1 of the lifecycle is built to answer the first question. The second is a liquidity question, and before 2022 many banks and clearing houses were not asking it with anything like the same rigour.

The UK case ran the same mechanism through a different market and compressed it into days. Liability-driven investment strategies used by UK defined-benefit pension schemes employ leveraged gilt exposure — typically through repo and interest-rate swaps — to match long-dated liabilities. When gilt yields rose sharply in September 2022, those positions generated collateral calls. Meeting the calls required selling assets, and the most liquid assets to sell were gilts, which pushed yields higher still, which generated further calls. That is a self-reinforcing loop, not a market move, and it was arrested by the Bank of England intervening in the gilt market. The pension schemes were not insolvent and their liabilities were in fact improving as yields rose. They were caught by the timing mismatch between a collateral obligation that is immediate and an asset position that is not.

Both episodes are the same failure at Stage 8, and the failure is procyclicality. Collateral requirements rise precisely when collateral is hardest to obtain, because the volatility that drives initial margin models upward is the same volatility that moves variation margin, and it hits every participant simultaneously. The system that converts credit risk into liquidity risk does so for everyone at once. That is not a flaw in any single institution's collateral framework; it is a property of the framework's design, and it can only be managed by holding liquidity against it.

What changed is a genuine addition to the discipline rather than a tightening of an existing control. Banks and their clients now stress test collateral requirements as a liquidity exercise: what would a severe move in the relevant risk factor demand in cash, over what horizon, in which currency, and could the counterparty produce it? Credit assessment of derivative counterparties increasingly includes an explicit view of margin liquidity capacity — the size and quality of the unencumbered asset pool relative to plausible calls — alongside the traditional view of solvency. Clearing houses came under sustained scrutiny for the procyclicality of their margin models, and the anti-procyclicality tools in those models moved from a technical annex to a supervisory priority. And the eligibility question — what a counterparty may post, and how quickly non-cash collateral can be transformed into the cash a variation margin call actually requires — became a first-order design choice rather than an operational detail.

Case Failure type Stage it maps to The lesson
Archegos, 2021 Exposure measurement and limits 3, 4 and 7 Margin that does not re-rate is not margin; concentration you cannot see is leverage you cannot limit
Lehman, 2008 Legal and operational 2 and 10 Netting is worth exactly what the documentation says it is worth
Margin crisis, 2022 Collateral and liquidity 8 The ability to post is a different question from the ability to pay

Five technologies are changing this business, and each attaches to a specific stage

Technology in counterparty risk is not a general modernisation story; each of the five trends below plugs into one identifiable point in the ten-stage lifecycle and does one identifiable thing there. That framing is useful precisely because it is unforgiving. A change that does not attach to a stage is not changing the business, and the section closes with a filter for telling the two apart.

AI-assisted credit analysis and early warning attaches to Stage 1

Credit assessment has always been constrained by how much a human analyst can read. A credit officer covering a portfolio of two hundred names cannot read every filing, every earnings transcript, every trade-press item and every covenant document for all of them, so coverage is triaged by size and by known concern — which means deterioration in a name nobody was worried about is found late. Machine reading removes the constraint on volume. Models can process filings, news flow, supply-chain signals and market data continuously across the whole portfolio and surface changes for human attention, which converts credit review from a periodic exercise into a monitored one. This is developed in detail in the next section.

Real-time and intraday exposure computation attaches to Stage 7

The overnight batch described in Part 2 exists because the computation used to be too slow to run more than once a day, and that constraint is dissolving. A full Monte Carlo revaluation of a bank's whole derivative portfolio across thousands of paths and a hundred time steps was, for years, an all-night job — which is why the exposure a trader sees at eleven in the morning is a number calculated from the previous evening's close, adjusted approximately. Faster hardware and better algorithms make intraday and, for some portfolios, near-real-time recomputation feasible. The consequence is not merely fresher reporting. It changes what a limit can be: a limit checked against a stale number is a soft constraint, while a limit checked against a live one is a hard one, and the Archegos case is an argument for the difference.

RegTech automation of SA-CCR, FRTB-CVA and EMIR reporting attaches to Stages 9 and 6

Regulatory calculation and reporting is high-volume, rule-driven, heavily audited work, which makes it the natural home of automation. SA-CCR requires a defined sequence of hedging-set construction, supervisory factors and aggregation for every netting set; FRTB-CVA under BA-CVA or SA-CVA requires its own sensitivity computations; EMIR REFIT requires reporting in ISO 20022 XML with a defined field set and matching obligations on both sides of the trade. Automating these does not reduce the capital charge — the charge is what the rule says it is — but it reduces the error rate, shortens the close, and makes reconciliation between the risk view and the regulatory view something that happens continuously rather than at quarter end.

Cloud and GPU compute for XVA attaches to Stages 5 and 9

The economics of XVA changed when the sensitivities became cheap. A CVA number is one calculation; hedging it requires the derivatives of that number with respect to every credit spread, rate and FX point in the model, which by brute-force revaluation means thousands of full simulations. Adjoint algorithmic differentiation computes the whole gradient at a small multiple of the cost of a single valuation, and elastic GPU capacity lets a bank rent an enormous amount of compute for the hours it is needed rather than owning it year-round. Together they take a calculation that was run monthly and make it a daily or intraday one. That is what makes an XVA desk able to hedge rather than merely report.

Straight-through collateral processing and automated dispute resolution attaches to Stage 8

Collateral operations was for a long time a business of spreadsheets, email and telephone calls, and the 2022 episode showed what happens when volumes spike against that infrastructure. Straight-through processing means the call is calculated, issued, agreed, instructed and settled without manual intervention, with electronic messaging replacing email and a tri-party agent or a central platform handling the movement. Automated reconciliation compares portfolios trade by trade between the two counterparties before a dispute arises, so that the mismatch is identified as a booking difference or a valuation difference rather than as a number nobody can explain. The effect is on ageing more than on volume: disputes still occur, but they are diagnosed in hours.

Trend Attaches to Why it matters What it does not solve
AI credit analysis and early warning Stage 1 Continuous monitoring across the whole portfolio, not just watched names The decision to act; a flag is not a downgrade
Real-time exposure computation Stage 7 Turns limits from soft to hard constraints Model quality; a fast wrong number is still wrong
RegTech for SA-CCR, FRTB-CVA, EMIR Stages 6 and 9 Fewer errors, faster close, continuous reconciliation The size of the capital charge itself
Cloud and GPU with AAD Stages 5 and 9 Makes daily XVA sensitivities affordable, so hedging is possible Hedge availability; no compute creates a CDS on Helvetia
Straight-through collateral and dispute automation Stage 8 Disputes diagnosed in hours, calls settled same day Whether the counterparty has the cash to post

Apply one filter to every technology proposal in this business: does it change the exposure, the capital or the loss — or only the reporting of them?

A change that moves the exposure is one that alters what the bank is actually carrying: dynamic margining that raises the requirement as a portfolio concentrates, or intraday recomputation that stops a limit being breached in the first place. A change that moves the capital alters the regulatory or economic charge: an IMM approval, a netting set restructuring, an SA-CVA implementation that recognises hedges the basic approach ignores. A change that moves the loss alters what happens on the bad day: a close-out process that is rehearsed, a dispute backlog that is cleared, documentation that is current.

Everything else is reporting. Reporting has real value — a dashboard that surfaces an aged dispute has done something useful — but it should be funded and justified as reporting, at reporting prices, rather than presented as risk reduction. The uncomfortable version of this test is that a great many large programmes in this discipline, when held against it, turn out to have improved the description of the risk while leaving the risk exactly where it was.

Generative AI earns its place where it changes exposure or capital, and nowhere else

The honest starting position is that counterparty risk is an unusually good fit for language models and an unusually hostile environment for them. Good, because the discipline is full of long documents that nobody has time to read carefully — filings, credit agreements, ISDA Schedules, CSAs, dispute correspondence — and extracting structured facts from unstructured text is exactly what these models do well. Hostile, because the outputs are capital-relevant, supervised and number-critical, and because a system that cannot explain itself has no place in a regulated model inventory. Both things are true at once, and the practical answer is not enthusiasm or refusal but a careful match of use case to control.

Credit analysis and early warning: reading everything, so a human can read the right thing

The constraint on credit monitoring has never been analytical skill; it has been reading capacity. A credit officer responsible for a portfolio of corporates covers the largest and the most worrying names properly and the rest on a schedule. Between scheduled reviews, deterioration is detected by whatever happens to reach the officer's desk — a rating action, a news story, a salesperson's remark. Language models change the arithmetic: annual reports, quarterly filings, earnings call transcripts, regulatory announcements, trade press and covenant documentation can be read continuously across every name, with the model summarising what has changed and flagging language that suggests stress.

What this produces is a triage layer, and the distinction matters for governance. The model does not assign the internal rating; it tells the credit officer which of two hundred names deserves attention this week and why, with the source passages attached. For a counterparty like Helvetia — a mid-size industrial with no public CDS, whose proxy curve comes from a sector index and therefore lags name-specific news badly — this is a genuine improvement in the bank's information position, because the proxy curve will not move on a covenant amendment or a plant closure but a model reading the filings will. The output is a hypothesis for a human to test, which is the correct role.

The risk to control here is the illusion of coverage. A flagging system that reads everything creates a strong impression that nothing has been missed, and it will miss things — particularly anything that is not written down anywhere it can read. It also generates false positives at a rate that erodes attention if the threshold is set carelessly, which is the classic failure of every early-warning system ever built. The metric that matters is not how many flags are raised but what proportion of flags led to an action and what proportion of downgrades were preceded by one.

Legal document extraction: the underrated win, because these parameters are capital-relevant

An ISDA Schedule and a CSA between them contain perhaps thirty parameters that drive the bank's exposure calculation, and in most institutions those parameters were keyed in by hand from a PDF. The threshold below which no collateral is called. The minimum transfer amount. The independent amount, if any. The list of eligible collateral and the haircut applied to each type. The call frequency and the notification deadlines. Downgrade triggers, which change the threshold or require an additional amount if the counterparty's rating falls. The governing law, the netting provisions, the additional termination events. Every one of these is an input to the collateralised exposure simulation, and therefore to the CVA charge and to the SA-CCR or IMM capital number.

Extraction at scale is a large win precisely because the alternative is so bad. Manual key-in of a legal parameter into a risk system is unverified, undated, done once, and rarely revisited when the agreement is amended. A model that reads the executed document, extracts each parameter with a citation to the clause it came from, and presents it for confirmation against what the system currently holds does three things: it populates new agreements faster, it reconciles the existing inventory against the underlying documents, and it produces an auditable link between the contract and the risk parameter that has never previously existed. Banks running this exercise on legacy portfolios routinely find discrepancies — thresholds that do not match, eligible collateral schedules that were never updated after an amendment, downgrade triggers that were missed entirely.

The reason to call this the underrated use case is that it fixes a Lehman-shaped problem with a tool that did not exist in 2008. Documentation hygiene became a discipline after Lehman because the industry learned what it costs when the paperwork does not match the system. It has remained a manual, expensive and therefore partially completed discipline ever since. A capability that can read a Schedule and tell you what your system should say is a direct attack on the failure mode that Part 2's Stage 2 exists to prevent.

Collateral dispute investigation: explaining why two banks disagree

A margin dispute is a reconciliation problem wrapped in a communication problem. Two institutions have valued the same portfolio and produced different numbers. The difference may be a trade one side has and the other does not, a trade booked with different terms on the two sides, a different curve or fixing, a different treatment of an option's early exercise, or a collateral balance that has not settled. Resolving it means comparing two trade populations, identifying the divergences, ranking them by contribution to the gap, and then explaining the result to a counterparty in language they can act on.

This work is well suited to a model because it is largely interpretive rather than computational. The comparison itself is arithmetic and the systems already do it; what takes the time is working out which of four hundred small differences actually explains the gap, and writing the explanation. A model that ingests both portfolios, the valuation breakdowns and the dispute correspondence, then produces a ranked explanation with supporting numbers, compresses days of analyst work into minutes. The benefit lands on dispute ageing, which the previous section identified as one of the sharpest indicators of operational quality — and an aged dispute is an uncollateralised exposure that nobody has decided to own.

Code and model documentation for the model inventory

A smaller but real use is the documentation burden that surrounds every model in the inventory. Model documentation must describe the methodology, the implementation, the calibration, the assumptions and the known limitations, and it must be kept current as the model changes — a task that competes for time with building the model and usually loses. Models that can read the implementation and produce a first draft of the technical documentation, or that can flag where documentation has diverged from code, reduce a real drag on the quantitative teams. This is a productivity gain rather than a risk reduction, and it should be claimed as one; the validation of the model remains entirely a human exercise, for reasons the next paragraphs make plain.

The three risks that govern all of it

The first is confidentiality, and in this specific context it is a wall problem rather than a data-protection problem. Counterparty-risk work touches watchlists, internal ratings, credit committee papers and deteriorating-name reviews — material that is price-sensitive and that sits behind an information barrier for exactly that reason. A model that is prompted with watchlist content, or that has been trained or fine-tuned on it, and is then accessible to users on the public side of the wall, is a breach of that barrier regardless of whether anyone intended it. This is not solved by a policy statement. It is solved by architecture: segregated deployments, retrieval scoped to the entitlements of the individual user, no training on restricted material, and logging that can demonstrate after the fact who saw what.

The second is hallucination in a context where every number is consequential. A model that summarises a filing and invents a covenant ratio, or that extracts a threshold of EUR 10m from a document that says EUR 1m, has not made a stylistic error — it has fed a wrong parameter into a capital calculation. The mitigation is structural rather than exhortative: every extracted figure carries a citation to the source passage, extraction is confirmed against the document by a human before it reaches a production system, and the model is never the system of record. Where a number can be checked deterministically, it is checked deterministically. The model proposes; the arithmetic disposes.

The third binds hardest, and it is explainability and model risk governance. A counterparty-risk model is a regulated model. It sits in an inventory, it has an owner and a validation status, and a supervisor is entitled to ask how it works, why it produces the number it produces, and what would happen if an assumption were wrong. A large language model is difficult to interrogate in those terms — it is not reproducible in the way a Monte Carlo engine with a fixed seed is reproducible, its behaviour changes when the underlying model version changes, and its reasoning cannot be inspected in a form that satisfies a validation function. That does not disqualify it. It determines where in the chain it may sit. A model that drafts a credit summary for an officer to accept or reject is a productivity tool operating behind a human decision. A model that sets a parameter which flows into an exposure calculation without confirmation is a model in the regulatory sense, and it must be inventoried, validated, version-controlled and monitored like any other.

Use of GenAI The gain The governing risk
Credit analysis and early warning Continuous coverage of the whole portfolio, not just watched names Confidentiality and MNPI; false-positive fatigue
ISDA and CSA parameter extraction Capital-relevant parameters populated and reconciled with citations Hallucination in a number-critical field; needs human confirmation
Dispute investigation Ranked explanation of valuation gaps; shorter dispute ageing Confidentiality of counterparty data; explanation must be verifiable
Model and code documentation Removes a real drag on quantitative teams Documentation that describes intent rather than the code as built
Anything feeding a production number Speed at scale Full model risk governance: inventory, validation, versioning

  1. A bank has had no counterparty losses for ten years. What does this tell you about the quality of its counterparty risk management?

  2. Which of the six risks are priced, which are structured around, and which are operated against?

  3. Archegos, Lehman and 2022 map onto which stages, and why are those three cases not a random selection?

  4. A utility with an investment-grade rating and rising profits cannot meet a margin call. Explain.

  5. Why can a generative AI model draft a credit summary but not set a CSA threshold in the risk system unconfirmed?

  6. What is the single filter to apply to any technology proposal in this discipline?

  7. Why does a collateralised portfolio still generate a CVA charge?

  8. What does reverse stress testing do that ordinary stress testing does not?

  9. Why is compression the only lever in this module that reduces capital without reducing business?

5

  1. No counterparty losses in ten years. What does it tell you?

  2. Which risks are priced into the client charge?

  3. Which case maps to collateral and liquidity?

The six ideas that matter most

Five parts, ten stages, one running example and a great deal of vocabulary reduce to six claims. Each is stated as a claim rather than a topic, because each is arguable and each has consequences for how a bank is organised, what it charges and where it loses money.

Counterparty risk is a derivative, and a derivative must be priced

The charge a bank adds for counterparty risk is not a fee, a margin or a provision — it is the price of an option the bank has written. When the bank faces Helvetia on an uncollateralised ten-year cross-currency swap, it has implicitly sold protection on Helvetia's default, contingent on the swap being in the money at the time. That is a compound option: a payoff that depends on a credit event and on a market level simultaneously. CVA is its value. The EUR 4.2m charge at inception in the illustrative Helvetia case — 91 basis points of the EUR 460m notional, roughly 11 basis points per annum running — is what that option is worth on day one, and Figure 5 and the valuation triangulation in Part 2's Stage 5 show exactly how the expected positive exposure, the proxy credit curve, the loss given default and the discounting combine to produce it.

Because it is a price, it behaves like one, and everything that follows is a consequence of that. It changes every day with rates, FX and credit spreads. It has sensitivities that can be computed and hedged. It has a profit and loss, which someone must own — which is why Part 1 placed the XVA desk on the trading floor rather than in the risk function. It is an accounting number under IFRS 13, not merely a risk number, which means it moves reported earnings. And when the spread on the name moves, the loss is realised without any default: the year-three illustration in which Helvetia is downgraded to BB+, the proxy spread widens to 480 basis points and the reserve rises from EUR 4.2m to EUR 9.6m is a EUR 5.4m charge to this quarter's earnings on a counterparty that has paid every cash flow on time.

The claim has an organisational edge to it. A bank that treats counterparty risk as a control function will produce reports; a bank that treats it as a price will produce a hedged book and a charge that clients pay. The difference shows up in whether the second-line veto described in Part 1 is the only lever the institution has, or whether the risk can be transferred internally, held on a book, and managed as a position. The statistic behind Basel III's CVA capital charge — roughly two-thirds of counterparty-credit losses in 2007 to 2009 came from CVA mark-to-market moves rather than actual defaults — is the empirical case for the claim, and it is why the framework was rebuilt around it.

Exposure is a distribution, not a number, and every control is a choice of which point on it to act on

There is no such thing as "the exposure" to a derivative counterparty; there is a probability distribution of possible exposures for every future date, and each control in this business picks a different point on it. Expected exposure is the average on a given date. Expected positive exposure is the average across time and is what feeds the price. Potential future exposure is a high percentile — commonly 97.5 per cent — and is what limits are set against. Effective EPE is the regulatory variant that feeds capital under the internal model method. Figure 5 in Part 2 draws all of this for the Helvetia swap: an expected positive exposure of EUR 22.4m against a peak potential future exposure of EUR 68m at year 4.4, both illustrative, against a group limit line at EUR 75m.

The gap between those two numbers is the whole point. EUR 22.4m and EUR 68m describe the same trade on the same day. One is the right input to a price, because a price should reflect what the bank expects to lose. The other is the right input to a limit, because a limit exists to defend against a bad outcome rather than a typical one. Using one where the other belongs is not a rounding error; it either overcharges every client the bank has or leaves it undefended in the tail. That is why the pre-deal check at Stage 4 measures incremental potential future exposure against the limit while the pricing at Stage 5 uses expected positive exposure, and why a reader who confuses the two will misunderstand every number in this module.

The consequence for anyone building or reviewing this machinery is a single question worth asking of any figure on any report: which point on the distribution is this, and what is it for? A great many arguments in this discipline dissolve the moment that question is answered, because the parties turn out to have been discussing different percentiles of the same simulation. And because these are model outputs rather than observed prices, the question has a second half — how confident are we that the distribution is right? — which is where backtesting and model validation enter, and why model risk sits in the six-risk taxonomy alongside default itself.

The netting set, not the trade, is the unit of risk

A trade has a mark-to-market; only a netting set has an exposure. A netting set is the group of trades that can be legally collapsed into a single net claim if the counterparty defaults, which is defined by an enforceable close-out netting provision in the ISDA Master Agreement covering those trades with that legal entity in that jurisdiction. Inside the set, positive and negative marks offset. Outside it, they do not. The illustrative whole-portfolio cascade used in Part 2's Stage 8 makes the size of the effect visible: gross positive mark-to-market of EUR 142m collapses to EUR 38m after close-out netting, then to EUR 12m after variation margin, then to approximately zero after initial margin.

Nearly every operational and system consequence in this module descends from that fact. The pre-deal check at Stage 4 is not asking what a trade is worth but what it does to the netting set, which is why the answer can be that a large trade reduces exposure and consumes no limit at all. The exposure simulation at Stage 7 prices along paths and aggregates into netting sets, because that is the level at which the number means anything. SA-CCR is calculated per netting set. Collateral is called per netting agreement. And Part 4's lineage diagram exists because the systems must agree on which trades belong to which set — a reference-data question that looks clerical and is not.

The claim also explains why documentation is a risk discipline rather than a legal formality. The netting set is a legal construct before it is a computational one, and its boundaries are exactly the boundaries the netting opinion supports. Lehman is the demonstration: 900,000 transactions across 6,500 entities, where the reduction from gross to net was realised wherever the paperwork was clean and evaporated wherever it was not. A netting set that a court does not recognise was never a netting set; it was a hopeful assumption sitting inside a capital calculation.

Collateral converts credit risk into liquidity risk; it does not delete risk

A collateral agreement changes the question from "will they pay?" to "can they post today?", and those are genuinely different questions with different answers. The cascade above shows what collateral achieves — EUR 38m of netted exposure reduced to EUR 12m by variation margin and to approximately zero by initial margin — and it is a large achievement. What it does not do is remove risk from the system. It replaces a credit exposure that materialises rarely and slowly with a liquidity obligation that materialises frequently and immediately, on both sides of the trade, and it does so for every participant at the same moment, because the volatility that drives the calls is common.

Two residues survive collateralisation and both matter. The first is the margin period of risk: ten business days uncleared, five cleared, during which the mark-to-market moves and no collateral arrives. That is why initial margin exists and why a fully collateralised book still carries a CVA charge, smaller but not zero. The second is the liquidity residue itself, which the 2022 margin crisis converted from a theoretical point into tens of billions of euros of emergency government liquidity lines for European utilities and a Bank of England intervention in the gilt market. Neither the utilities nor the pension schemes were failing credits. They were solvent institutions that could not produce cash on the morning it was demanded.

The practical instruction is to run the collateral machine as a liquidity function with a credit purpose, not as a credit function with a liquidity side-effect. That means stress testing what a severe move would demand in cash, in which currency and over what horizon; assessing a counterparty's margin liquidity capacity alongside its solvency at Stage 1; treating eligibility and collateral transformation as design choices rather than operational details; and watching dispute ageing and call timeliness as leading indicators. Helvetia, as an NFC− below the EMIR clearing threshold with no CSA at all, is the sharpest teaching case precisely because it sits at the other extreme: no collateral, no liquidity obligation, and the full credit exposure retained and priced.

The loss happens at close-out, but it is decided at documentation

Stage 10 is where money is lost and Stage 2 is where the amount was determined, and the years between them are when the determination could have been checked and was not. A close-out requires the bank to identify every trade with the correct legal entity, value them on a defensible basis, net where netting is enforceable, apply the collateral it actually holds and can actually realise, serve notices within contractual deadlines, and then defend all of it. Each of those depends on something agreed long before: the netting provision, the netting opinion, the security interest, the segregation arrangement, the rehypothecation right, the definition of the close-out amount, the additional termination events.

Lehman is the evidence and the numbers are the argument. Close-out netting worked where documentation was clean and failed where it was not, and the valuation litigation ran for more than a decade because determining a close-out amount in a dislocated market is a judgement the determining party must later justify with evidence. The failures were not exotic. They were unexecuted amendments, trades booked against the wrong entity, collateral in accounts whose segregation status was not what the agreement implied, opinions with qualifications nobody had reread. On any ordinary day these are administration. On one particular day they are the difference between a net claim and a gross one.

This is the claim that most often gets nodded at and least often gets funded, and the asymmetry is worth naming. Documentation hygiene has no revenue attached, produces no visible improvement in any daily metric, and competes for investment against things that do. Its entire value is contingent on an event that may not occur for a decade. That is also precisely why it is the natural home of the generative AI use case set out above: reading a Schedule and reconciling its parameters against the risk system is exactly the kind of large, dull, high-value verification that was previously too expensive to do at scale, and it attacks the failure mode that this claim identifies.

Real transformation moves the exposure or the capital — including GenAI, which is governed as tightly as it is useful

The filter proposed earlier in this part is the right one to end on: does a change alter the exposure, the capital or the loss — or only the reporting of them? Dynamic margining on a synthetic prime book alters the exposure and would have altered Archegos. Intraday recomputation turns a limit from an advisory number into a binding one. Adjoint algorithmic differentiation on rented GPU capacity makes daily XVA sensitivities affordable, which is what allows an XVA desk to hedge rather than merely report — a change in the loss profile, not the report. An SA-CVA implementation that recognises hedges the basic approach ignores alters the capital. A dashboard, however good, alters the description.

The purest example of the filter is not a technology at all. Portfolio compression terminates economically redundant trades and leaves the net risk position untouched, which releases capital and operational cost without removing a single unit of risk the bank was actually carrying. It required no model, no vendor and no rule mandating it; it happened because SA-CCR and the leverage ratio charge on the contract population while the bank runs the net position, and someone noticed. It is worth holding up as the standard against which transformation proposals should be judged, because it changes a capital number by a large multiple of what most technology programmes achieve, and it changes no report at all.

Generative AI passes the filter in three specific places and fails it in many others. Reading filings, news and covenant documents across the whole portfolio changes what the bank knows before a downgrade, and therefore what exposure it chooses to carry. Extracting thresholds, minimum transfer amounts, eligible collateral and downgrade triggers from ISDA Schedules and CSAs changes parameters that feed the collateralised exposure simulation and therefore capital — the underrated win, because it attacks the Lehman failure mode with a tool that did not exist in 2008. Explaining a collateral dispute compresses ageing, and an aged dispute is uncollateralised exposure nobody has decided to own. Those three change something real.

And the governance tightens in exact proportion to the value. Confidentiality is an information-barrier problem here, not merely a data-protection one, because watchlists and credit reviews are price-sensitive. Hallucination in a field where every output is a number is not a stylistic defect but a wrong parameter in a capital calculation. And explainability binds hardest of all, because a counterparty-risk model is a regulated model with an owner, a validation status and a supervisor entitled to ask how it works. The human in the loop is what makes the capability usable rather than what limits it — it converts an output nobody can interrogate into a decision a named person made and can defend. That is the same bargain the whole discipline runs on: the model produces the number, and a person is accountable for it.

Process Cheat Sheet — The Whole Lifecycle on One Page

A single reference for the entire counterparty-risk lifecycle. The ten stages in order, the numbers worth memorising, and the discipline in one sentence.

The ten stages at a glance

Stage Timing Core activity Key output Lead and systems
1 Onboarding and credit assessment 2–12 weeks KYC, financial analysis, internal rating, appetite decision Internal rating and PD; approved product list Credit risk; client onboarding, rating and reference-data systems
2 Legal architecture 4–16 weeks Negotiate ISDA Master, Schedule, CSA; obtain netting opinion Enforceable netting set; collateral terms Legal and documentation; contract repository, legal-terms database
3 The limit framework Annual, reviewed continuously Set PFE, tenor, settlement and concentration limits Limit structure by measure and tenor Credit risk committee; limit management system
4 The pre-deal check Seconds Incremental PFE against remaining limit headroom Approve, refuse or escalate Sales and trading; pre-deal check engine, limit system
5 XVA pricing and the charge Minutes Compute CVA, FVA, KVA; quote the all-in price The XVA charge embedded in the price XVA desk; pricing library, exposure and credit-curve data
6 Booking and trade capture T+0 Capture terms, allocate to netting set and legal entity Golden-source trade record; confirmation Middle office; trade capture, confirmation, reporting feeds
7 The daily exposure cycle Every night, 18:00–06:00 Curve build, path generation, pricing, netting-set aggregation EE, EPE, EEPE, PFE profiles; limit utilisation Risk engine team; Monte Carlo grid, market data, batch scheduler
8 Collateral and margin Daily, intraday in stress Calculate, issue, agree and settle VM and IM calls Collateral balances; disputes; residual exposure Collateral operations; collateral system, tri-party agent, custodian
9 XVA hedging and capital Continuous Hedge CVA sensitivities; compute SA-CCR or IMM EAD and RWA Hedged XVA book; capital and RWA numbers XVA desk and capital team; sensitivity engine, capital calculator
10 Distress, default and close-out Days to years Watchlist, termination notice, close-out amount, netting, recovery Net claim; realised loss; litigation record Workout, legal, operations; contract repository, valuation evidence

The numbers worth memorising

Metric Rule of thumb
SA-CCR alpha 1.4 — EAD = 1.4 × (RC + PFE add-on)
MPoR, uncleared 10 business days; extended where disputes or illiquidity apply
MPoR, cleared 5 business days
PFE confidence level 95–99%, commonly 97.5% for limit setting
Monte Carlo paths 2,000–10,000 for production exposure runs; more for tail measures
Time grid Dense in the near term, sparse at long tenors; typically 50–150 dates
UMR final phase threshold AANA above EUR 8bn brings a counterparty into scope for initial margin
CVA rule of thumb CVA ≈ EPE × credit spread × duration (a sanity check, not a valuation)
LGD convention 60%, i.e. a 40% recovery assumption for senior unsecured
Hazard rate approximation Hazard rate ≈ credit spread / LGD; a 300bp spread at 60% LGD implies ~5% annual default intensity
Survival probability Approximately exp(−hazard rate × time) under a flat curve
FX supervisory factor, SA-CCR 4% of adjusted notional
Basel III CVA rationale Roughly two-thirds of 2007–09 counterparty-credit losses came from CVA mark-to-market, not defaults
Capital ratio used in this module 10.5% applied to RWA (illustrative working assumption)

The Helvetia case in numbers — illustrative only

Item Value (illustrative)
Trade 10-year fixed-for-fixed EUR/USD cross-currency swap, USD 500m against EUR 460m
Counterparty BBB− equivalent, NFC− below the EMIR clearing threshold, no CSA
Proxy credit spread 180bp at 5 years, 220bp at 10 years; LGD 60%
EPE EUR 22.4m
Peak PFE at 97.5% EUR 68m at year 4.4, against a EUR 75m group limit (91% utilisation)
CVA at inception EUR 4.2m = 91bp of EUR notional ≈ 11bp per annum running
FVA / KVA / total XVA EUR 1.1m / EUR 0.9m / EUR 6.2m
SA-CCR add-on and EAD Add-on EUR 18.4m; EAD = 1.4 × (0 + 18.4) = EUR 25.8m
RWA and capital RWA EUR 25.8m at 100% risk weight; capital at 10.5% = EUR 2.7m
Portfolio cascade Gross EUR 142m → netted EUR 38m → after VM EUR 12m → after IM ~zero
Year-3 stress Swap at +EUR 41m, counterparty at BB+, spread 480bp, CVA reserve EUR 9.6m — a EUR 5.4m realised loss with no default
Typical compression reduction 60-80% of gross notional and trade count in a mature bilateral book
Stressed EPE Capital takes the higher of stressed and current calibration
Large exposures ceiling 25% of Tier 1 capital per group of connected clients; tighter between G-SIBs
Measure by purpose Real-world for PFE and limits; risk-neutral for CVA and XVA pricing
Nightly cycle window Roughly 18:00 to 06:00, with pricing and aggregation the critical path

The discipline in one sentence

Counterparty risk is the only risk a bank takes without being paid a spread for it, so it must be measured as an exposure, priced as a derivative, funded as a balance sheet item, and hedged as a market position — and every failure in this business is a failure to do one of those four things.